PrudentBench v1.1 · updated 20 October 2026

Which AI model is suitable for Dutch legal research?

PrudentBench does not run the external benchmarks itself. We summarise published results, add one screening of our own on Dutch cases and test every model against a compliance gate.

As of October 2026 Claude Opus 5, Claude Fable 5.1 and Muse Spark 1.3 (Max) tie for the highest score on the Vals AI Legal Research Bench: 55.29 percent all-pass [1]; Gemini 4 Argon follows at 54.81 percent and has been generally available in the Europe multi-region of Google Cloud since 13 October [55]. After the compliance test, processing in the EU, a European contracting party and no training on customer data by contract, thirteen of the twenty-three models remain. Six models are open weights and five of those are listed by a European open-weights host [45], but the public endpoint there does not fix the processing location [44]; they wait for a dedicated EU endpoint.

Highest legal-research score: 55.29 %Highest legal-research scoreVals AI, measured 19 October 2026 [1]
Models through the compliance gate: 13 of 23Models through the compliance gatePrudentBench v1.1
External benchmarks for legal work: 2External benchmarks for legal workVals AI Legal Research Bench and Harvey LAB (Vals) [1][2]
Models assessed, 6 open weights: 23Models assessed, 6 open weightsPrudentBench v1.1

Beau Jonkhout, Chief Technology Officer at Prudai · Last updated 20 October 2026 · PrudentBench v1.1

1. The scorecard

Twenty-three models, six of them open weights, two benchmarks and one compliance gate. Only models that pass the gate receive a PrudentScore.

Passes the compliance gateRejected by the compliance gate

Legal research: all-pass in percentVals AI Legal Research Bench, all-pass, measured 19 October 2026 [1]. Hatched bars are rejected by the compliance gate.

Quality against cost per taskCost per completed task in dollars, logarithmic, against all-pass on legal research [1]. Top left is cheap and good.

Passes the compliance gateRejected by the compliance gate

PrudentScore v1.1

Only models that pass the gate get a score, on a scale from 0 to 100.

1
Gemini 4 Argon80points
2
Claude Opus 572points
3
Claude Opus 5.563points
4
Gemini 3.8 Flash62points
5
Claude Sonnet 5.562points
6
Claude Sonnet 559points
7
GPT-6.1 Sol59points
8
GPT-5.6 Sol58points
9
GPT-5.6 Luna56points
10
GPT-6 Luna54points
11
GPT-5.6 Terra53points
12
GPT-6 Sol46points
13
Mistral Medium 3.523points

All figures as a table

PrudentBench v1.1 scorecard: twenty-three AI models on legal research, legal agent tasks, cost per task and the compliance gate.
Model	Provider and route	Legal research	Legal agent tasks	Cost per task	Compliance gate	PrudentScore
Gemini 4 Argon	Google via Vertex AI (Europe multi-region)	54.81%	19.58%	$6.45	Passes	80
Claude Opus 5	Anthropic via Google Cloud Vertex AI (EU regions)	55.29%	6.67%	$6.76	Passes	72
Claude Opus 5.5	Anthropic via Google Cloud Vertex AI (Europe multi-region)	50.48%max effort, with fallback on refusals [49]	3.75%	$25.16	Passes	63
Gemini 3.8 Flash	Google via Vertex AI (EU regions)	38.94%	10.00%	$1.63	Passes	62
Claude Sonnet 5.5	Anthropic via Google Cloud Vertex AI (Europe multi-region)	48.08%max effort, with fallback on refusals [54]	2.92%	$14.83	Passes	62
Claude Sonnet 5	Anthropic via Google Cloud Vertex AI (EU regions)	41.83%	5.00%	$2.72	Passes	59
GPT-6.1 Sol	OpenAI via Microsoft Azure OpenAI (EU Data Zone)	38.46%	5.42%	$3.38	Passes	59
GPT-5.6 Sol	OpenAI via Microsoft Azure OpenAI (Sweden Central)	48.08%	2.50%	$21.61	Passes	58
GPT-5.6 Luna	OpenAI via Microsoft Azure OpenAI (EU Data Zone)	36.54%	1.25%	$0.85	Passes	56
GPT-6 Luna	OpenAI via Microsoft Azure OpenAI (EU Data Zone)	30.29%	2.92%	$0.44	Passes	54
GPT-5.6 Terra	OpenAI via Microsoft Azure OpenAI (EU Data Zone)	41.35%	0.83%	$7.91	Passes	53
GPT-6 Sol	OpenAI via Microsoft Azure OpenAI (EU Data Zone)	28.85%	1.67%	$4.90	Passes	46
Mistral Medium 3.5	Mistral AI (Paris), also via the Azure EU Data Zone	9.14%	0.42%	$1.31	Passes	23
Claude Fable 5.1A mandatory 30-day retention period without zero data retention, and on Vertex AI data sharing with the model maker is required.	Anthropic, via Google Cloud	55.29%	6.67%	$23.06	Rejected	no score
Muse Spark 1.3 (Max)Available only through Meta's own API; no route via a hyperscaler and no published guarantee of EU processing or an EU contracting party. The cheap contributor tier trains on customer traffic.	Meta	55.29%without Max: 40.87	23.75%	$0.59	Rejected	no score
GLM 5.3On the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.	Z.ai (open weights), through a European open-weights host	49.04%	8.33%	$2.24	Rejected	no score
Grok 4.6On Azure available only as Global Standard, without an EU Data Zone [22][23].	xAI via Microsoft Azure	48.08%	15.83%	$1.53	Rejected	no score
Kimi K3On the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.	Moonshot AI (open weights), through a European open-weights host	46.15%	12.92%	$3.47	Rejected	no score
GLM 5.3 FlashOn the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.	Z.ai (open weights), through a European open-weights host	45.19%	6.67%	$0.12	Rejected	no score
DeepSeek V4.1 FlashOn the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.	DeepSeek (open weights), through a European open-weights host	41.35%	6.67%	$0.25	Rejected	no score
DeepSeek V4 Pro (0813)On the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.	DeepSeek (open weights), through a European open-weights host	40.87%	7.50%	$1.09	Rejected	no score
GPT-6 AstraOn Azure only Global Standard and a US Data Zone; no EU Data Zone [23].	OpenAI via Microsoft Azure	39.42%	5.42%	$10.58	Rejected	no score
Mistral Large 4Public preview: under Mistral’s commercial terms Mistral may use customer data and outputs of preview models to train its models, and an opt-out or zero data retention does not apply there [59].	Mistral AI, public preview	31.73%	15.83%	$5.33	Rejected	no score

Legal research: Vals AI Legal Research Bench, all-pass in percent [1]. Legal agent tasks: Harvey LAB in the Vals implementation, all-pass in percent [2]. Cost per task: Vals AI, in US dollars [1]. Where a benchmark publishes an effort setting, it is part of the model name: Muse Spark 1.3 (Max) is a different setting from plain Muse Spark 1.3, which reaches 40.87 percent. Vals ran Claude Opus 5.5 at max effort, with another Claude model as a fallback on refusals; on the Legal Research Bench that did not change the score [49]. Vals also ran Claude Sonnet 5.5 at max effort, with Claude Sonnet 5 as a fallback on refusals; on the Legal Research Bench it stepped in on just over 5 percent of the questions [54]. Vals ran Gemini 4 Argon at high effort [51].

Benchmark figures: the Vals AI Legal Research Bench [1] and the Vals implementation of Harvey LAB [2], measured 19 October 2026 (Legal Research Bench) and 18 October 2026 (Vals implementation of Harvey LAB, version 1), accessed 20 October 2026. The PrudentScore is a composition by Prudai B.V. and is not a figure produced by the benchmark operators.

2. Why yes, why no: model by model

What each model does well, where it breaks down and what we consider it suitable for. The separate line about the Intelligence Index carries its source, because it comes from a different measurement than the table above.

Gemini 4 Argon

Google via Vertex AI (Europe multi-region)

PassesPrudentScore 80

54.81 percent all-pass, 0.48 points below the shared top [1].
After Muse Spark (Max) the best on legal agent tasks: 19.58 percent [2].
6.45 dollars per task, about four times Gemini 3.8 Flash [1].
Has not yet been through our own screening on Dutch cases.

Suitable for: Deep case analysis and case-law research, and legal tasks in several steps.

Artificial Analysis Intelligence Index v4.3.2: 52.6 at high effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

54.81 percent all-pass on the Vals AI Legal Research Bench, 0.48 percentage points below the three models tied at 55.29 percent, and more than 15 points above Gemini 3.8 Flash [1]. Vals ran the model at high effort [51].
After Muse Spark 1.3 (Max) the best result in this list on legal agent tasks: 19.58 percent all-pass on the Vals implementation of Harvey LAB [2].
Generally available on Google Cloud since 13 October 2026, with availability and processing in the Europe multi-region [55].

Weak

6.45 dollars per task in the Vals measurement, about four times Gemini 3.8 Flash; Vals prices at 4 and 20 dollars per million tokens, the price from 1 December 2026. In the Europe multi-region Google charges 10 percent more [1][51][56].
No explicit context cache: only the automatic discount on repeated context from 6,144 tokens [55].
Has not yet been through our own screening on Dutch cases.

Claude Opus 5

Anthropic via Google Cloud Vertex AI (EU regions)

PassesPrudentScore 72

Tied highest legal-research score: 55.29 percent all-pass [1].
The most robust model in our screening over long, source-heavy conversations.
6.76 dollars per task, more than four times Gemini 3.8 Flash [1].
Low on legal agent tasks: 6.67 percent all-pass [2].

Suitable for: Deep case analysis, case-law research and memos in which many sources come together.

Artificial Analysis Intelligence Index v4.3.2: 50.8 at max effort; Artificial Analysis now marks the model as deprecated. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Joint highest result on the Vals AI Legal Research Bench: 55.29 percent all-pass, tied with Claude Fable 5.1 and Muse Spark 1.3 (Max), measured 19 October 2026 [1].
In our own screening the most dependable model in long conversations that call many sources.
Discounted repeated context keeps case-wide work affordable per unit of work.

Weak

6.76 dollars per task in the Vals measurement, more than four times Gemini 3.8 Flash [1].
A deep analysis takes noticeably longer than a quick question.
It lags on legal agent tasks: 6.67 percent all-pass on the Vals implementation of Harvey LAB [2].

Claude Opus 5.5

Anthropic via Google Cloud Vertex AI (Europe multi-region)

PassesPrudentScore 63

50.48 percent all-pass, fifth in this list [1].
Not a Covered Model; zero data retention available [14][48].
25.16 dollars per task, the most expensive in this list [1].
Weak on legal agent tasks: 3.75 percent all-pass [2].

Suitable for: Legal research and case analysis where quality weighs more than cost per task.

Artificial Analysis Intelligence Index v4.3.2: 57.6 at max effort with fallback, the highest of these twenty-three and of the whole index. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

50.48 percent all-pass on the Vals AI Legal Research Bench, fifth in this list, after the three models tied at 55.29 percent and Gemini 4 Argon [1]. Vals ran the model at max effort, with Claude Opus 5 and Claude Opus 4.8 as a fallback on refusals; on this benchmark that did not change the score [49].
Generally available on Google Cloud in the Europe multi-region, both for model availability and for processing [47].
Not a Covered Model: according to the provider it is available with zero data retention, like earlier Opus models [14][48].

Weak

25.16 dollars per task in the Vals measurement, the most expensive task in this list and more than three and a half times Claude Opus 5, which scores higher [1].
Weak on legal agent tasks: 3.75 percent all-pass on the Vals implementation of Harvey LAB, below Claude Opus 5 [2].
Has not yet been through our own screening on Dutch cases.

Gemini 3.8 Flash

Google via Vertex AI (EU regions)

PassesPrudentScore 62

1.63 dollars per task, about a quarter of Claude Opus 5 [1].
After Gemini Argon the best on agent tasks among those that pass the gate: 10.00 percent [2][20].
38.94 percent all-pass, below Claude Sonnet 5 and Claude Opus 5 [1].
Calls fewer sources than Claude Sonnet 5 on the same question.

Suitable for: Short questions, summaries and first explorations where the answer is checked afterwards.

Artificial Analysis Intelligence Index v4.3.2: 40.9 at high effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

1.63 dollars per task in the Vals measurement, about a quarter of Claude Opus 5 [1].
In our own screening no demonstrable quality difference with Claude Sonnet 5 on short questions, at about 40 percent lower cost per task [1].
After Gemini 4 Argon the best result on legal agent tasks among the models that pass the gate: 10.00 percent all-pass on the Vals implementation of Harvey LAB [2][20].

Weak

38.94 percent all-pass on legal research, below Claude Sonnet 5 and well below Claude Opus 5 [1].
It waits noticeably longer before the first character once sources are called.
It calls fewer sources than Claude Sonnet 5 on the same question.

Claude Sonnet 5.5

Anthropic via Google Cloud Vertex AI (Europe multi-region)

PassesPrudentScore 62

48.08 percent all-pass, more than six points above Claude Sonnet 5 [1].
Not a Covered Model; zero data retention available [14][53].
14.83 dollars per task, more than twice Claude Opus 5 [1].
Weak on legal agent tasks: 2.92 percent all-pass [2].

Suitable for: Legal research where the higher score than Claude Sonnet 5 is worth the higher cost per task; Claude Opus 5 scores higher in the Vals measurement at a lower cost per task [1].

Artificial Analysis Intelligence Index v4.3.2: 56.0 at max effort with fallback, after Claude Opus 5.5 the highest of these twenty-three. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

48.08 percent all-pass on the Vals AI Legal Research Bench, more than six percentage points above Claude Sonnet 5 [1]. Vals ran the model at max effort, with Claude Sonnet 5 as a fallback on refusals; it stepped in on just over 5 percent of the questions [54].
Generally available on Google Cloud in the Europe multi-region, both for model availability and for processing [52].
Not a Covered Model: according to the provider it is available with zero data retention, like Claude Opus 5.5 and Claude Sonnet 5 [14][53].

Weak

14.83 dollars per task in the Vals measurement, more than five times Claude Sonnet 5 and more than twice Claude Opus 5, which scores higher [1].
Weak on legal agent tasks: 2.92 percent all-pass on the Vals implementation of Harvey LAB, below Claude Sonnet 5 [2].
Has not yet been through our own screening on Dutch cases.

Claude Sonnet 5

Anthropic via Google Cloud Vertex AI (EU regions)

PassesPrudentScore 59

Starts answering almost immediately, which suits follow-up questions.
41.83 percent all-pass on legal research, above Gemini 3.8 Flash [1].
Pricier per task than Gemini 3.8 Flash: 2.72 against 1.63 dollars [1].
Low on legal agent tasks: 5.00 percent all-pass [2].

Suitable for: Conversations with follow-up questions, and work where speed and source calls have to go together.

Artificial Analysis Intelligence Index v4.3.2: 38.2 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

It starts answering almost immediately, which suits conversations with follow-up questions.
It works well with source calls and keeps track across several steps.
41.83 percent all-pass on the Vals AI Legal Research Bench, above Gemini 3.8 Flash [1].

Weak

More expensive per completed task than Gemini 3.8 Flash: 2.72 against 1.63 dollars in the Vals measurement [1].
Clearly lower on legal research than Claude Opus 5.
Low on legal agent tasks: 5.00 percent all-pass on the Vals implementation of Harvey LAB, half of Gemini 3.8 Flash [2].

GPT-6.1 Sol

OpenAI via Microsoft Azure OpenAI (EU Data Zone)

PassesPrudentScore 59

Azure EU Data Zone with Microsoft Ireland as the contracting party [23].
3.38 dollars per task, about a sixth of GPT-5.6 Sol [1].
38.46 percent all-pass, almost ten points below GPT-5.6 Sol [1].
Has not yet been through our own screening on Dutch cases.

Suitable for: Legal research within the OpenAI family in the EU Data Zone at a low cost per task; the highest OpenAI score on legal research still belongs to GPT-5.6 Sol [1].

Artificial Analysis Intelligence Index v4.3.2: 51.8 at max effort, the highest of the OpenAI models that pass the gate. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-6.1-sol is listed by name in Microsoft's Europe table, in its 29 September 2026 version [21][23].
3.38 dollars per task in the Vals measurement, about a sixth of GPT-5.6 Sol [1].
Clearly higher on legal research than GPT-6 Sol: 38.46 against 28.85 percent all-pass [1].

Weak

On legal research still almost ten percentage points below GPT-5.6 Sol [1].
Low on legal agent tasks: 5.42 percent all-pass on the Vals implementation of Harvey LAB [2].
Has not yet been through our own screening on Dutch cases.

GPT-5.6 Sol

OpenAI via Microsoft Azure OpenAI (Sweden Central)

PassesPrudentScore 58

48.08 percent all-pass, the best of the OpenAI models here [1].
Azure EU Data Zone with Microsoft Ireland as contracting party [23].
21.61 dollars per task, the priciest after Opus 5.5 and Fable 5.1 [1].
Low on legal agent tasks: 2.50 percent all-pass [2].

Suitable for: Legal research that calls for a model family other than Claude or Gemini, for example as an independent jury model, and where cost per task weighs less.

Artificial Analysis Intelligence Index v4.3.2: 47.0 at max effort; Artificial Analysis now marks the model as deprecated. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

48.08 percent all-pass on the Vals AI Legal Research Bench, the best result of the seven OpenAI models in this list [1].
Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-5.6-sol is listed by name in Microsoft's Europe table [23].
A third model family alongside Claude and Gemini, which makes it usable as an independent jury model.

Weak

21.61 dollars per task in the Vals measurement, after Claude Opus 5.5 and Claude Fable 5.1 the most expensive model in this list [1].
A long wait before the first visible character.
Low on legal agent tasks: 2.50 percent all-pass on the Vals implementation of Harvey LAB [2].

GPT-5.6 Luna

OpenAI via Microsoft Azure OpenAI (EU Data Zone)

PassesPrudentScore 56

36.54 percent all-pass for 0.85 dollars per task [1].
Azure EU Data Zone with Microsoft Ireland as contracting party [23].
Low on legal agent tasks: 1.25 percent all-pass [2].
Has not yet been through our own screening on Dutch cases.

Suitable for: Short questions and summaries where cost per task weighs heavily; slightly stronger on legal research than GPT-6 Luna [1].

Artificial Analysis Intelligence Index v4.3.2: 37.3 at max effort; Artificial Analysis now marks the model as deprecated. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

36.54 percent all-pass on legal research for 0.85 dollars per task, above its successor GPT-6 Luna [1].
Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-5.6-luna is listed by name in Microsoft's Europe table [21][23].

Weak

On legal research almost 19 percentage points below Claude Opus 5 [1].
Low on legal agent tasks: 1.25 percent all-pass on the Vals implementation of Harvey LAB [2].
Has not yet been through our own screening on Dutch cases.

GPT-6 Luna

OpenAI via Microsoft Azure OpenAI (EU Data Zone)

PassesPrudentScore 54

0.44 dollars per task, the cheapest that passes the gate [1].
Azure EU Data Zone with Microsoft Ireland as contracting party [23].
30.29 percent all-pass, below its predecessor GPT-5.6 Luna [1].
Has not yet been through our own screening on Dutch cases.

Suitable for: Many short, simple tasks where cost per task weighs heavily and the answer is checked afterwards.

Artificial Analysis Intelligence Index v4.3.2: 38.1 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

0.44 dollars per task in the Vals measurement: the cheapest task among the models that pass the gate [1].
Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-6-luna is listed by name in Microsoft's Europe table [21][23].

Weak

Low on legal research: 30.29 percent all-pass, below its predecessor GPT-5.6 Luna [1].
Low on legal agent tasks: 2.92 percent all-pass on the Vals implementation of Harvey LAB [2].
Has not yet been through our own screening on Dutch cases.

GPT-5.6 Terra

OpenAI via Microsoft Azure OpenAI (EU Data Zone)

PassesPrudentScore 53

41.35 percent all-pass, after GPT-5.6 Sol the best OpenAI result [1].
Azure EU Data Zone with Microsoft Ireland as contracting party [23].
7.91 dollars per task, more than Claude Opus 5 [1].
Second lowest on legal agent tasks: 0.83 percent [2].

Suitable for: Legal research within the OpenAI family for just over a third of the cost per task of GPT-5.6 Sol, at a lower score [1].

Artificial Analysis Intelligence Index v4.3.2: 42.1 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

41.35 percent all-pass on legal research, after GPT-5.6 Sol the best result of the OpenAI models in this list [1].
Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-5.6-terra is listed by name in Microsoft's Europe table [21][23].

Weak

7.91 dollars per task in the Vals measurement, more than Claude Opus 5, which scores higher [1].
The second lowest of these twenty-three on legal agent tasks: 0.83 percent all-pass [2].
Has not yet been through our own screening on Dutch cases.

GPT-6 Sol

OpenAI via Microsoft Azure OpenAI (EU Data Zone)

PassesPrudentScore 46

In the Azure EU Data Zone, unlike GPT-6 Astra [23].
4.90 dollars per task, less than a quarter of GPT-5.6 Sol [1].
28.85 percent all-pass, the lowest of the OpenAI models here [1].
Has not yet been through our own screening on Dutch cases.

Suitable for: Not much any more: on the same EU Data Zone route GPT-6.1 Sol scores higher on legal research and on legal agent tasks, at a lower cost per task [1][2].

Artificial Analysis Intelligence Index v4.3.2: 47.6 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Processing in the Azure EU Data Zone with Microsoft Ireland as the contracting party: gpt-6-sol is listed by name in Microsoft's Europe table, in its 22 September 2026 version [21][23].
A GPT-6 generation model that, unlike GPT-6 Astra, is available in the EU Data Zone [23].
4.90 dollars per task in the Vals measurement, less than a quarter of GPT-5.6 Sol [1].

Weak

Low on legal research: 28.85 percent all-pass, the lowest result of the OpenAI models in this list and far below GPT-5.6 Sol [1].
Low on legal agent tasks: 1.67 percent all-pass on the Vals implementation of Harvey LAB [2].
Has not yet been through our own screening on Dutch cases.

Mistral Medium 3.5

Mistral AI (Paris), also via the Azure EU Data Zone

PassesOpen weightsPrudentScore 23

The only European model provider in this list, with a European contracting party.
Open weights, so suitable for fully self-hosted deployment.
9.14 percent all-pass on legal research, the lowest in this list [1].
Artificial Analysis Intelligence Index 14.2, the lowest here; measured since v4.3.2 [37].

Suitable for: A candidate for fully self-hosted deployment and for narrow tasks, not for broad legal research.

Artificial Analysis Intelligence Index v4.3.2: 14.2, since v4.3.2 a measurement of its own instead of an estimate. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

The only European model provider in this list, with a European contracting party.
Open weights, so it is suitable for fully self-hosted deployment.
Available through the Azure EU Data Zone under Microsoft's processing agreement; mistral-medium-3-5 is listed by name in Microsoft's Europe table [21][23], and Mistral offers a European endpoint of its own [32][33].

Weak

9.14 percent all-pass on the Vals AI Legal Research Bench, by far the lowest result in this list [1].
0.42 percent all-pass on legal agent tasks, likewise the lowest of these twenty-three [2].
Since Intelligence Index v4.3.2 Artificial Analysis has a measurement of its own for this model, 14.2, the lowest in this list; the PrudentScore now rests on all four components [37].

Claude Fable 5.1

Anthropic, via Google Cloud

Rejected

Gate closed: A mandatory 30-day retention period without zero data retention, and on Vertex AI data sharing with the model maker is required.

Tied highest legal-research score: 55.29 percent all-pass [1].
Same model family as Claude Opus 5, so prompts carry over.
23.06 dollars per task, more than three times Claude Opus 5 [1].
A mandatory 30-day retention window, no zero data retention [13][16].

Suitable for: Not for work with client data: the compliance gate closes on retention and on the required data sharing.

Artificial Analysis Intelligence Index v4.3.2: 53.4 at max effort with fallback, the third highest of these twenty-three, after Claude Opus 5.5 and Claude Sonnet 5.5. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Joint highest result on the Vals AI Legal Research Bench: 55.29 percent all-pass, level with Claude Opus 5 and Muse Spark 1.3 (Max) [1].
The same result on legal agent tasks as Claude Opus 5: 6.67 percent all-pass [2].
The same model family as Claude Opus 5, so prompts and working methods transfer.

Weak

23.06 dollars per task in the Vals measurement, more than three times Claude Opus 5 at the same score [1].
The provider has designated it a Covered Model since 31 August 2026, with a mandatory 30-day retention period and without zero data retention [13][14][16].
On Google Cloud the Fable line requires data sharing with the model maker to be switched on for abuse monitoring [17].

Muse Spark 1.3 (Max)

Strong

Joint highest result on the Vals AI Legal Research Bench: 55.29 percent all-pass at Max effort, level with Claude Opus 5 and Claude Fable 5.1 [1].
The best result on legal agent tasks in this list: 23.75 percent all-pass on the Vals implementation of Harvey LAB [2].
0.59 dollars per completed task in the Vals measurement, a fraction of Claude Opus 5 at the same score [1].

Weak

The 55.29 percent applies only to the Max effort setting, which the provider's launch post still described as being in safety testing; the version anyone can buy scores 40.87 percent [1][28].
Available only through Meta's own API, not via Azure, Google Cloud or Bedrock, and without a published guarantee of EU processing or a European contracting party [29].
The cheapest price tier, the contributor tier, is used to improve the provider's products and therefore trains on customer traffic [29].

GLM 5.3

Z.ai (open weights), through a European open-weights host

RejectedOpen weightsAt European open-weights host

Gate closed: On the public endpoint of the European open-weights host the region reads Global, so EU processing is not contractually guaranteed [44]. The gate only opens on a dedicated EU endpoint or self-hosted infrastructure, where the region is fixed by contract.

Highest open-weights score on legal research: 49.04 percent all-pass [1].
Open weights (GLM-5.3 licence), listed by the European open-weights host [40][45].
Public endpoint of the host: region Global, location not guaranteed [44].
Has not been through our own screening on Dutch cases.

Suitable for: A candidate for a dedicated EU endpoint or self-hosting; not assessed in our own screening.

Artificial Analysis Intelligence Index v4.3.2: 44.8 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

49.04 percent all-pass on the Vals AI Legal Research Bench, the highest open-weights result in this list and above GPT-5.6 Sol and Grok 4.6 [1].
Open weights under the GLM-5.3 licence, 753 billion parameters, commercial use allowed with restrictions [40].
Listed in the model catalogue of the European open-weights host, so it can be tried without a GPU fleet of our own [45].

Weak

On the public endpoint the processing location is not guaranteed [44].
8.33 percent all-pass on legal agent tasks, at 2.24 dollars per task in the Vals measurement [1][2].
Has not been through our own screening on Dutch cases.

Grok 4.6

xAI via Microsoft Azure

Rejected

Gate closed: On Azure available only as Global Standard, without an EU Data Zone [22][23].

48.08 percent all-pass, level with GPT-5.6 Sol and Sonnet 5.5 [1].
Strong on legal agent tasks: 15.83 percent all-pass [2][26].
On Azure only Global Standard, without an EU Data Zone [22][23].
Almost a minute of silence before the first token in our test.

Suitable for: Not for work with client data while there is no EU Data Zone. We will reassess it once there is one.

Artificial Analysis Intelligence Index v4.3.2: 44.3 at high effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

48.08 percent all-pass on the Vals AI Legal Research Bench, level with GPT-5.6 Sol and Claude Sonnet 5.5 [1].
Through Azure, Microsoft is the processor and the contracting party, not the model maker: prompts and completions are not available to the providers of models sold by Azure [21].
Strong on legal agent tasks: 15.83 percent all-pass on the Vals implementation of Harvey LAB [2][26].

Weak

On Azure available only as Global Standard, so without a guaranteed EU region; grok-4.6 is absent from Microsoft's Europe table [22][23].
In our own test almost a minute of silence before the first character, and no discount on repeated context.

Kimi K3

Moonshot AI (open weights), through a European open-weights host

RejectedOpen weightsAt European open-weights host

Open weights, so it can run on our own infrastructure or a dedicated EU endpoint [44].
46.15 percent all-pass, above Claude Sonnet 5 and Gemini 3.8 Flash [1].
Public endpoint of the host: region Global, location not guaranteed [44].
Too slow and too erratic for interactive use in our own August 2026 test.

Suitable for: Work on self-hosted infrastructure or on a dedicated EU endpoint; less suited to interactive use.

Artificial Analysis Intelligence Index v4.3.2: 43.6 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Open weights, so it can run on our own infrastructure or on a dedicated EU endpoint without the model maker in the data path [44].
46.15 percent all-pass on the Vals AI Legal Research Bench, above Claude Sonnet 5 and Gemini 3.8 Flash [1].
Low cost per task in the Vals measurement: 3.47 dollars, at 12.92 percent all-pass on legal agent tasks [1][2].

Weak

On the public endpoint the processing location is not guaranteed; only a dedicated endpoint fixes the region contractually [44].
In our own August 2026 test too slow and too erratic for interactive use, and without a discount on repeated context more expensive per unit of work than Claude Opus 5.

GLM 5.3 Flash

Z.ai (open weights), through a European open-weights host

RejectedOpen weightsAt European open-weights host

45.19 percent all-pass for 0.12 dollars per task, the cheapest in this list [1].
Open weights under MIT, 18 billion active parameters [41].
Public endpoint of the host: region Global, location not guaranteed [44].
Has not been through our own screening on Dutch cases.

Suitable for: A candidate for a dedicated EU endpoint or self-hosting; not assessed in our own screening.

Artificial Analysis Intelligence Index v4.3.2: 41.8. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

45.19 percent all-pass on legal research for 0.12 dollars per task: the cheapest task in this entire list [1].
Open weights under MIT, 320 billion parameters of which 18 billion active, so light enough for self-hosting [41].
Listed in the model catalogue of the European open-weights host [45].

Weak

On the public endpoint the processing location is not guaranteed [44].
6.67 percent all-pass on legal agent tasks [2].
Has not been through our own screening on Dutch cases.

DeepSeek V4.1 Flash

DeepSeek (open weights), through a European open-weights host

RejectedOpen weightsAt European open-weights host

41.35 percent all-pass for 0.25 dollars per task, just below Claude Sonnet 5 [1].
Open weights under MIT, listed by the European open-weights host [42][45].
Public endpoint of the host: region Global, location not guaranteed [44].
Has not been through our own screening on Dutch cases.

Suitable for: A candidate for a dedicated EU endpoint or self-hosting; not assessed in our own screening.

Artificial Analysis Intelligence Index v4.3.2: 39.5 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

41.35 percent all-pass on legal research for 0.25 dollars per task, just below Claude Sonnet 5 at a fraction of the cost [1].
Open weights under MIT, 552 billion parameters with 8 to 16 billion active per token [42].
Listed in the model catalogue of the European open-weights host [45].

Weak

On the public endpoint the processing location is not guaranteed [44].
6.67 percent all-pass on legal agent tasks [2].
Has not been through our own screening on Dutch cases.

DeepSeek V4 Pro (0813)

DeepSeek (open weights), through a European open-weights host

RejectedOpen weightsAt European open-weights host

40.87 percent all-pass for 1.09 dollars per task [1].
Open weights under MIT, released 13 August 2026 [43].
Public endpoint of the host: region Global, location not guaranteed [44].
Has not been through our own screening on Dutch cases.

Suitable for: A candidate for a dedicated EU endpoint or self-hosting; not assessed in our own screening.

Artificial Analysis Intelligence Index v4.3.2: 36.0 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

40.87 percent all-pass on legal research for 1.09 dollars per task [1].
Open weights under MIT, 1.7 trillion parameters; the 13 August 2026 release replaces the earlier preview [43].
Listed in the model catalogue of the European open-weights host [45].

Weak

On the public endpoint the processing location is not guaranteed [44].
7.50 percent all-pass on legal agent tasks [2].
Has not been through our own screening on Dutch cases.

GPT-6 Astra

OpenAI via Microsoft Azure

Rejected

Gate closed: On Azure only Global Standard and a US Data Zone; no EU Data Zone [23].

Strong at agentic work and at programming tasks.
After Opus 5.5, Sonnet 5.5 and Fable 5.1 the highest Artificial Analysis index here [37].
Low on legal research: 39.42 percent all-pass [1].
No EU Data Zone on Azure, only Global Standard and US [23].

Suitable for: Not for work with client data while there is no EU Data Zone. We will reopen the assessment once there is one and the legal-research score comes up.

Artificial Analysis Intelligence Index v4.3.2: 52.7 at max effort. Source: Artificial Analysis (artificialanalysis.ai), accessed 1 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

Strong on agentic work and on programming tasks.
High general intelligence: after Claude Opus 5.5, Claude Sonnet 5.5 and Claude Fable 5.1 the highest score in this list on the Artificial Analysis Intelligence Index [37].
Once Azure ships an EU Data Zone, the route to this model is already in place.

Weak

Low on legal research: 39.42 percent all-pass on the Vals AI Legal Research Bench [1].
On Azure only Global Standard and a US Data Zone are available, no EU Data Zone, which keeps the compliance gate closed [23].

Mistral Large 4

Mistral AI, public preview

Rejected

Gate closed: Public preview: under Mistral’s commercial terms Mistral may use customer data and outputs of preview models to train its models, and an opt-out or zero data retention does not apply there [59].

15.83 percent all-pass on legal agent tasks, level with Grok 4.6 [2].
European provider; according to Mistral the preview is served from its own European data centres [58].
Public preview: Mistral may train on customer data, even with zero data retention [59].
31.73 percent all-pass on legal research [1].

Suitable for: Not for work with client data while it is a preview. We will assess it again at general availability or once the open weights are out.

Artificial Analysis Intelligence Index v4.3.2: 38.4, measured as Mistral Large 4 Preview. Source: Artificial Analysis (artificialanalysis.ai), accessed 20 October 2026 [37]. artificialanalysis.ai

More about this model

Strong

On legal agent tasks 15.83 percent all-pass on the Vals implementation of Harvey LAB, level with Grok 4.6 [2].
According to Mistral trained and served as a preview in Mistral’s own European data centres, with an announced European deployment that Mistral operates itself; we have not checked that endpoint ourselves [58].
Mistral announces open weights for the end of October 2026 [58].

Weak

31.73 percent all-pass on the Vals AI Legal Research Bench, the lowest result among the models that do not pass the gate [1][57].
5.33 dollars per task in the Vals measurement, more than three times Gemini 3.8 Flash [1].
As long as the model is a preview, Mistral may train on the data entered, even with zero data retention switched on [59].

3. Compliance matrix

The gate only opens when all three requirements are covered: processing in the EU, a contracting party in the EU or EEA, and a contractual exclusion of training on customer data. Where a provider publishes no figure, the cell reads "see provider" and carries no estimate.

Compliance matrix: processor, contracting party, EU data processing, training on customer data and retention per model.
Model	EU processing	EU contracting party	No training	Gate
Gemini 4 Argon	yes	yes	yes	Passes
Claude Opus 5	yes	yes	yes	Passes
Claude Opus 5.5	yes	yes	yes	Passes
Gemini 3.8 Flash	yes	yes	yes	Passes
Claude Sonnet 5.5	yes	yes	yes	Passes
Claude Sonnet 5	yes	yes	yes	Passes
GPT-6.1 Sol	yes	yes	yes	Passes
GPT-5.6 Sol	yes	yes	yes	Passes
GPT-5.6 Luna	yes	yes	yes	Passes
GPT-6 Luna	yes	yes	yes	Passes
GPT-5.6 Terra	yes	yes	yes	Passes
GPT-6 Sol	yes	yes	yes	Passes
Mistral Medium 3.5	yes	yes	yes	Passes
Claude Fable 5.1	no	yes (Google Cloud)	yes	Rejected
Muse Spark 1.3 (Max)	no	no	partly	Rejected
GLM 5.3	partly	yes	yes	Rejected
Grok 4.6	no	yes	yes	Rejected
Kimi K3	partly	yes	yes	Rejected
GLM 5.3 Flash	partly	yes	yes	Rejected
DeepSeek V4.1 Flash	partly	yes	yes	Rejected
DeepSeek V4 Pro (0813)	partly	yes	yes	Rejected
GPT-6 Astra	no	yes	yes	Rejected
Mistral Large 4	partly	yes (Mistral AI SAS)	no	Rejected

Why the gate closes

Claude Fable 5.1, EU processing: On Google Cloud only with mandatory data sharing with the model maker for abuse monitoring and a mandatory 30-day retention window at the vendor [15][17]
Muse Spark 1.3 (Max), EU processing: No published guarantee of EU processing [29][30]
Muse Spark 1.3 (Max), EU contracting party: Meta, no published EU/EEA entity
Muse Spark 1.3 (Max), No training: Yes on the contributor tier; no on the standard tier [29]
GLM 5.3, EU processing: No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]
Grok 4.6, EU processing: No, Global Standard only, no EU Data Zone [22][23]
Kimi K3, EU processing: No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]
GLM 5.3 Flash, EU processing: No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]
DeepSeek V4.1 Flash, EU processing: No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]
DeepSeek V4 Pro (0813), EU processing: No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]
GPT-6 Astra, EU processing: No, Global Standard and US Data Zone only [23]
Mistral Large 4, EU processing: European deployment announced, not checked by us [58]
Mistral Large 4, No training: Yes, allowed for preview models (terms §4.2 and §4.3) [59]

Processor, contracting party and retention as a table

Compliance matrix: processor, contracting party, EU data processing, training on customer data and retention per model.
Model	Processor	Contracting party	EU data processing	Training on customer data	Retention
Gemini 4 Argon	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [55]	No, contractually excluded [38]	Abuse logging only after a classifier signal, up to 90 days in the chosen multi-region; not a Google Frontier Model [17]
Claude Opus 5	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [18][19]	No, contractually excluded [38]	0 (zero data retention) [15][16]
Claude Opus 5.5	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [47]	No, contractually excluded [38]	0 (zero data retention) [15][16]; not a Covered Model [14]
Gemini 3.8 Flash	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [18][19]	No, contractually excluded [38]	0 (zero data retention) [15][16]
Claude Sonnet 5.5	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [52]	No, contractually excluded [38]	0 (zero data retention) [15][16]; not a Covered Model [14]
Claude Sonnet 5	Google Cloud	Google Cloud EMEA Limited	Vertex AI, Europe multi-region [18][19]	No, contractually excluded [38]	0 (zero data retention) [15][16]
GPT-6.1 Sol	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, EU Data Zone (including Sweden Central) [21][23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
GPT-5.6 Sol	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, Sweden Central (EU Data Zone) [23]	No, contractually excluded [21]	Contractual zero data retention; abuse monitoring with reviewers inside the EEA [21]
GPT-5.6 Luna	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, EU Data Zone (including Sweden Central) [21][23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
GPT-6 Luna	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, EU Data Zone (including Sweden Central) [21][23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
GPT-5.6 Terra	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, EU Data Zone (including Sweden Central) [21][23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
GPT-6 Sol	Microsoft	Microsoft Ireland Operations Limited	Azure OpenAI, EU Data Zone (including Sweden Central) [21][23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
Mistral Medium 3.5	Mistral AI, or Microsoft when purchased through Azure	Mistral AI SAS (Paris), or Microsoft Ireland Operations Limited	European endpoint, or Azure EU Data Zone [23][32][33]	No, contractually excluded [21]	See provider
Claude Fable 5.1	Google Cloud	Google Cloud EMEA Limited	On Google Cloud only with mandatory data sharing with the model maker for abuse monitoring and a mandatory 30-day retention window at the vendor [15][17]	No, contractually excluded [38]	30 days mandatory, no zero data retention [13][16]
Muse Spark 1.3 (Max)	Meta	Meta, no published EU/EEA entity	No published guarantee of EU processing [29][30]	Yes on the contributor tier; no on the standard tier [29]	See provider
GLM 5.3	European open-weights host, public endpoint [45]	Established in the Netherlands [46]	No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]	No, excluded in the terms [44]	Zero data retention only when you switch it on yourself; otherwise kept for speculative decoding [44]
Grok 4.6	Microsoft (Models sold by Azure)	Microsoft Ireland Operations Limited	No, Global Standard only, no EU Data Zone [22][23]	No, contractually excluded [21]	Through Azure the Microsoft processing agreement applies [21]
Kimi K3	European open-weights host, public endpoint [45]	Established in the Netherlands [46]	No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]	No, excluded in the terms [44]	Zero data retention only when you switch it on yourself; otherwise kept for speculative decoding [44]
GLM 5.3 Flash	European open-weights host, public endpoint [45]	Established in the Netherlands [46]	No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]	No, excluded in the terms [44]	Zero data retention only when you switch it on yourself; otherwise kept for speculative decoding [44]
DeepSeek V4.1 Flash	European open-weights host, public endpoint [45]	Established in the Netherlands [46]	No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]	No, excluded in the terms [44]	Zero data retention only when you switch it on yourself; otherwise kept for speculative decoding [44]
DeepSeek V4 Pro (0813)	European open-weights host, public endpoint [45]	Established in the Netherlands [46]	No on the public endpoint: region Global, not guaranteed; yes on a dedicated endpoint (DPA §6.1) [44]	No, excluded in the terms [44]	Zero data retention only when you switch it on yourself; otherwise kept for speculative decoding [44]
GPT-6 Astra	Microsoft	Microsoft Ireland Operations Limited	No, Global Standard and US Data Zone only [23]	No, contractually excluded [21]	Abuse monitoring with reviewers inside the EEA [21]
Mistral Large 4	Mistral AI	Mistral AI SAS [59]	European deployment announced, not checked by us [58]	Yes, allowed for preview models (terms §4.2 and §4.3) [59]	Opt-out and zero data retention do not apply to preview models [59]

This matrix describes the route we would use, not every conceivable route to the same model. We never claim a provider by model family: Microsoft lists gpt-6.1-sol, gpt-6-sol, gpt-6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna and mistral-medium-3-5 by name in its Europe table, and gpt-6-astra and grok-4.6 not at all [23]. Our own subprocessors and processing agreement are documented in the trust center and in our privacy statement [39].

4. How we choose models

The scorecard shows how models perform on published benchmarks. Choosing one for a concrete application takes more: first the compliance gate, then six building blocks that weigh differently per task. Which models a Prudai product uses at any given moment is deliberately not on this page, so that the benchmark stands apart from our own choice.

First the gate

Processing in the EU, fixed by contract
A contracting party in the EU or EEA
A contractual exclusion of training on customer data

A model that fails any of these three is rejected, however high its score. The gate applies per route: the same model can pass through one provider and fail through another.

Then six building blocks

Quality on legal research
The Legal Research Bench as the heaviest signal: half of the PrudentScore.
Legal agent tasks
Work in several steps; every model still scores low here.
Cost per unit of work
Cost per completed task, only after the discount on repeated context.
Speed
Weighed per task: a question in a conversation is not a case analysis.
The EU route
The route decides the gate: party, region and contract.
Reliability is architecture
Citations checked word for word against the source text, independent of the model.

Cost per completed task, models that pass the gateVals AI, in US dollars per completed task, cheapest first [1].

Notes on the six building blocks

Quality on legal research: The Vals AI Legal Research Bench is the heaviest signal: lawyers write the questions and the grading rubrics there, and an answer only counts if it meets every rubric item; the grading itself is done by a language model, GPT-5.4. That is why it counts for half of the PrudentScore [1]. Per task we then look at what is needed: a summary asks something different from reconstructing conflicting authority.
Legal agent tasks: Work in several steps, with documents and tools, as the Vals implementation of Harvey LAB measures it. Every model still scores low there; the highest value on version 1 of the Vals leaderboard is 25.42 percent all-pass [2]. A high research score therefore says little about work carried out on its own.
Cost per unit of work: We compare cost per completed task, not per token, and only after the discount on repeated context: a lower list price means nothing while the provider gives no such discount. In this list the cost per task differs by more than a factor of a hundred [1].
Speed: A question in an ongoing conversation calls for a different model than a case analysis that may take minutes. Speed is not part of the PrudentScore; we weigh it per task, in our own screening.
The EU route: Not the model but the route decides whether the gate opens: which party processes, in which region and under which contract. A route only counts once the provider lists the model by name in its European offering; a model family is no proof [23].
Reliability is architecture: Citations are checked word for word against the retrieved source text. That check is independent of the model, so swapping models changes nothing about it.

5. Methodology and PrudentScore

The gate

A model is rejected, whatever its score, if any of these three is not covered:

no contractual processing in the EU
no contracting party in the EU or EEA
no contractual exclusion of training on customer data

PrudentScore v1.1

Only models that pass the gate get a score. The score adds four components and is rounded to whole points:

How the four components are weighted

50%Legal research
15%Legal agent tasks
15%General intelligence
20%Cost efficiency

The arithmetic and the results per model

50 percent legal research: Vals AI Legal Research Bench, all-pass, divided by the highest score of 55.29 [1].
15 percent legal agent tasks: Harvey LAB in the Vals implementation, all-pass, divided by the highest score of 25.42, the highest value on version 1 of the Vals leaderboard (Muse Spark 1.2, which is not in this list). Since 19 October 2026 Vals also publishes version 1.1, but it covers only part of this list; version 1 therefore remains the source [2].
15 percent general intelligence: Artificial Analysis Intelligence Index v4.3.2, divided by the highest score in the index, 57.6 for Claude Opus 5.5 at max effort with fallback [6][37].
20 percent cost efficiency: 1 minus a logarithmic scale of cost per task, with the cheapest and the most expensive task in this list as its ends: 0.12 dollars (GLM 5.3 Flash) as the floor and 25.16 dollars (Claude Opus 5.5) as the ceiling [1].

When a component is missing, the remaining weights are rescaled across what is left. The results below are computed by the page from this formula and rounded to whole points; they are stored nowhere as a number. A difference of a few points is not a meaningful difference: before rounding Claude Opus 5.5, Gemini 3.8 Flash and Claude Sonnet 5.5 reach 62.9, 62.0 and 61.8, rounded 63, 62 and 62. Read the order as an indication, not as the outcome of a significance test.

Results for v1.1, computed from this formula

Gemini 4 Argon4 of 4 components80
Claude Opus 54 of 4 components72
Claude Opus 5.54 of 4 components63
Gemini 3.8 Flash4 of 4 components62
Claude Sonnet 5.54 of 4 components62
Claude Sonnet 54 of 4 components59
GPT-6.1 Sol4 of 4 components59
GPT-5.6 Sol4 of 4 components58
GPT-5.6 Luna4 of 4 components56
GPT-6 Luna4 of 4 components54
GPT-5.6 Terra4 of 4 components53
GPT-6 Sol4 of 4 components46
Mistral Medium 3.54 of 4 components23

What we do ourselves

Our own assessment runs in four steps. First the compliance filter: a model rejected there is not measured any further. Then a signal wait: we wait until the first published results are stable rather than switching on release day. Next a screening on our own settings, so with our prompts, our source calls and our context length, not on the benchmark configuration. Only then a full benchmark.

Two things weigh heavily in that. Compliance is a gate, not a weighting factor: a model that fails the gate gets no second chance on quality. And the cache economy decides: a provider without a discount on repeated context is more expensive on case-wide work than the list price suggests.

Prudai is itself the supplier of LEO. That is why we cite figures only from parties that are not ours, publish the formula, also show the models that fail the gate with the reason, and keep our own model choice off this page.

What we do not measure

An independent, published benchmark for Dutch legal work does not exist yet. Our own Dutch measurement set (100 rulings × 10 questions) is being built; as soon as it passes our quality gate, we will publish the results here.

That gap can be checked against the literature. The only serious Dutch-language legal question-answering benchmark, bLLeQA, covers Belgian law, and its authors write that reliable Dutch resources are scarce [34]. The only benchmark that touches Dutch law, Multi-Legal-Bench, does metadata classification and outcome prediction on rulings, not legal research [35]. Dutch is present in the EU-wide knowledge benchmark EU-MMLU, but its only legal component is international law [36].

That set started as a screening on 20 Dutch cases with a blind jury and a check on case references. Until the extension is finished we publish no figures from it. The models added on 28 September and 1 October 2026 have not yet been through that screening.

Harvey LAB exists in three implementations, by Harvey itself [4][5], by Vals [2] and by Artificial Analysis [8], with outcomes that can differ by a factor of three. We consistently cite the Vals implementation on all-pass, because there every model runs on the same harness, and mention Artificial Analysis only in running text with attribution.

The PrudentScore is a composition by Prudai B.V. based on publicly published figures and does not represent the view or endorsement of Vals AI, Harvey or Artificial Analysis. Figures from Artificial Analysis appear in running text only, with attribution and a link, as that party's terms require [9].

6. What changed

Changes to this page, newest first.

20 October 2026
Gemini 4 Argon has been generally available on Google Cloud since 13 October 2026, with processing in the Europe multi-region [55], and now passes the compliance gate: thirteen of the twenty-three models pass. Under the existing formula PrudentScore v1.1 Gemini 4 Argon reaches 80 points, the highest in this list, above Claude Opus 5 (72). The scores of the other models do not change. Vals figures re-read, as at 19 October 2026 (Legal Research Bench): Kimi K3 rose to 46.15 percent all-pass at 3.47 dollars per task, the other models are unchanged [1]. Since 19 October 2026 Vals also publishes the Vals implementation of Harvey LAB, version 1.1, which covers only part of this list; this page therefore cites version 1, where Kimi K3 now reaches 12.92 percent [2]. Mistral Large 4 added as a rejected model: it is a public preview, and under Mistral’s terms Mistral may train on customer data there [59].
1 October 2026
Three models added: Gemini 4 Argon, Claude Sonnet 5.5 and GPT-6.1 Sol. Claude Sonnet 5.5 and GPT-6.1 Sol pass the compliance gate, so twelve of the twenty-two models now pass. Gemini 4 Argon lands just below the shared top at 54.81 percent all-pass, but is not yet generally available and therefore does not pass the gate yet [1][50]. Vals figures re-read as of 29 September (Legal Research Bench) and 30 September 2026 (Vals implementation of Harvey LAB): the cost per task of Claude Opus 5.5 rose to 25.16 dollars and that of Claude Sonnet 5 fell to 2.72 dollars [1]. Artificial Analysis adjusted GPT-6 Sol and GPT-6 Luna to 47.6 and 38.1 [37]. The formula remains PrudentScore v1.1; the ceiling of the cost scale follows the most expensive task to 25.16 dollars, so the existing PrudentScores shift slightly. Thanks to its lower cost per task Claude Sonnet 5 now reaches 59 points (was 57), above GPT-5.6 Sol; GPT-6 Sol reaches 46 (was 45).
28 September 2026
PrudentBench v1.1. Five models added: Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Luna and GPT-5.6 Terra. All five pass the compliance gate, so ten of the nineteen models now pass. The formula is now PrudentScore v1.1, because two of its ends are defined differently: the divisor for the Artificial Analysis Intelligence Index follows the highest score in that index to 57.6 (Claude Opus 5.5) [37], and the cost scale no longer sits between fixed amounts (0.59 and 23.06 dollars) but runs from the cheapest to the most expensive task in this list, now 0.12 to 23.61 dollars [1]. The existing PrudentScores shift as a result. Vals figures re-read as of 22 September 2026, unchanged for the existing models. The page no longer says which models Prudai's products use, not in this overview either: the benchmark stands apart from our own choice.
22 September 2026
Four open-weights models added (GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash, DeepSeek V4 Pro) with their availability at a European open-weights host. Kimi K3 and the new open-weights models no longer pass the gate: that host's public endpoint does not guarantee the processing location [44]. Artificial Analysis Intelligence Index updated to v4.3.2 (7 September 2026); the divisor in the formula moves to 53.4 accordingly [37]. Vals figures re-read as of 21 September 2026, unchanged.
6 September 2026
PrudentBench v1.0, first publication.

7. Frequently asked questions

The eleven questions we get most often about model choice and legal work.

What is the best AI model for legal advice?

On the Vals AI Legal Research Bench, a legal-research benchmark whose questions and grading rubrics are written by lawyers, Claude Opus 5, Claude Fable 5.1 and Muse Spark 1.3 (Max) share the lead at 55.29 percent all-pass, measured 19 October 2026 [1]. Gemini 4 Argon follows at 54.81 percent [1]. For a firm the compliance gate counts as well: processing in the EU, a European contracting party and no training on customer data. Of the three leaders, only Claude Opus 5 passes it; since its general availability on Google Cloud in the Europe multi-region Gemini 4 Argon passes it as well [55], and so has the highest PrudentScore in this list. The newer Claude Opus 5.5 passes the gate too, at 50.48 percent [1]. An independent, published benchmark for Dutch legal work does not exist yet.

Is ChatGPT GDPR-proof for law firms?

That does not depend on the model but on the route and the contract. A consumer subscription offers no processing agreement and no choice of processing region. Through Microsoft Azure OpenAI in the EU Data Zone, for example in Sweden Central, Microsoft Ireland is the contracting party, a processing agreement is in place and there is a contractual exclusion of training on customer data [21][23]. That is the route this page assesses for GPT models. Always assess your own situation and your own processing agreement.

May I enter client data into an AI model?

Only when the processing is covered: a processing agreement with the provider, no training on the data entered, a documented processing region and clear agreements on retention. Professional privilege does not disappear because a model sits in between, the responsibility stays with the firm. Our agreements are in the processing agreement at legal.prudai.com and our subprocessors are described in the trust center.

Claude or ChatGPT for legal research?

On the Vals AI Legal Research Bench, Claude Opus 5 reaches 55.29 percent all-pass and Claude Opus 5.5 50.48 percent, against 48.08 percent for GPT-5.6 Sol, the best result of the seven OpenAI models in this list, measured 19 October 2026 [1]. Cost per task varies widely: 6.76 dollars for Claude Opus 5, 21.61 for GPT-5.6 Sol and 25.16 for Claude Opus 5.5 [1]. On legal agent tasks all three score low: 6.67, 2.50 and 3.75 percent all-pass [2]. Through Google Cloud Vertex AI and through the EU Data Zone of Microsoft Azure OpenAI, all three pass the compliance gate.

Why does AI invent case law, and which model does it least?

A language model predicts text, and a plausible-looking case reference is as likely to the model as a real one. There is no public measurement of hallucination on Dutch case law, for any model [34][35], so we name no winner. The answer lies in architecture, not in model choice: LEO retrieves rulings at the source and checks every citation word for word against the retrieved text.

Which models process data in the EU?

Of the twenty-three models assessed, thirteen pass the gate: Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5.5 and Claude Sonnet 5 through Google Cloud Vertex AI in EU regions, Gemini 4 Argon and Gemini 3.8 Flash through the same route, GPT-6.1 Sol, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna through the EU Data Zone of Microsoft Azure OpenAI, and Mistral Medium 3.5 through a European endpoint [18][19][23][47][52][55]. Kimi K3, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro are open weights and listed by a European open-weights host [45], but on the public endpoint the processing location is not guaranteed; they only pass the gate on a dedicated EU endpoint or self-hosted infrastructure [44]. Claude Fable 5.1, Grok 4.6, GPT-6 Astra and Muse Spark are rejected [17][22][23][29], and so is Mistral Large 4 while it is a preview [59].

Are open-weights models automatically compliant?

No. Open weights say who may run the model, not where it runs. Six models in this list are open weights: Mistral Medium 3.5, Kimi K3, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro. The last five are listed by the European open-weights host we tested in August 2026 [45], but the public endpoint there shows Global as its region and does not fix the processing location; only a dedicated endpoint does so contractually [44]. Anyone who wants to use these models with client data therefore needs a dedicated EU endpoint or self-hosting. Once such a route is fixed by contract, we will reassess these models for this page, starting with GLM 5.3.

Does the provider train on my case files?

Not on the routes that pass the gate: Google Cloud and Microsoft contractually exclude training on customer data [21][38]. Watch the exceptions in the market. Claude Fable 5.1 carries a mandatory 30-day retention period without zero data retention [13][16], and the contributor tier of Muse Spark is used to improve the provider's products [29]. According to their provider, Claude Opus 5.5 and Claude Sonnet 5.5 are available with zero data retention [48][53].

What does the AI Act mean for my firm?

A firm that buys and uses an AI system is normally a deployer, not a provider. The obligations that come first are AI literacy among staff, transparency towards clients and human oversight of the outcome. Our role split, documentation and subprocessors are set out in the trust center at trust.prudai.com.

How often is PrudentBench updated?

Every quarter, and in between whenever a major model release or a change in a processing route calls for it. Every figure on this page carries its source and its measurement date. Figures older than six months we treat as expired: we verify them at the source again before republishing them.

Does Prudai measure itself, or only summarise?

Both, and we keep the two apart. The legal-research and agent-task scores come from Vals AI and Harvey LAB, benchmarks we do not run ourselves. In addition we carry out a screening of our own on Dutch cases, on our own settings rather than on the benchmark configuration. That screening is small and is being extended, and we publish no figures from it yet. The models added on 28 September and 1 October 2026 have not yet been through it.

8. Sources

The bracketed numbers in the text and in the tables refer to this list. Each source shows the date we last read it: the original sources on 6 September 2026, sources [40] to [46] on 22 September 2026, the sources on the models of 28 September and [47] to [49] on 28 September 2026, the index figures, Microsoft's Europe table, the list of Covered Models [14] and [50] to [54] on 1 October 2026, and the benchmark figures [1][2], the abuse-monitoring page [17] and [55] to [56] on 20 October 2026. The benchmark figures were measured on 19 October 2026 (Legal Research Bench) and 18 October 2026 (Vals implementation of Harvey LAB, version 1).

[1]Vals AI. Legal Research Bench. benchmark page, version 1, updated 19 October 2026.https://www.vals.ai/benchmarks/legal_researchAccessed 20 October 2026
[2]Vals AI. Harvey's Legal Agent Benchmark (HLAB leaderboard). leaderboard, version 1, archived, last updated 18 October 2026; since 19 October 2026 Vals shows version 1.1 by default, which covers part of this list.https://www.vals.ai/benchmarks/hlab-v1Accessed 20 October 2026
[3]Vals AI. Methodology. methodology page, undated.https://www.vals.ai/methodologyAccessed 6 September 2026
[4]Grupen, N., Pereyra, G. & Pereyra, J. (Harvey AI). Open-Sourcing Harvey's Long Horizon Legal Agent Benchmark. Harvey AI, 6 May 2026.https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmarkAccessed 6 September 2026
[5]Harvey AI. Harvey LAB: The Legal Agent Benchmark, v1.0. dataset and harness on GitHub, MIT licence.https://github.com/harveyai/harvey-labsAccessed 6 September 2026
[6]Artificial Analysis. Announcing the Artificial Analysis Intelligence Index v4.3. 7 September 2026; model pages show sub-version v4.3.2.https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3Accessed 22 September 2026
[7]Artificial Analysis. Artificial Analysis Intelligence Benchmarking Methodology. methodology page, undated.https://artificialanalysis.ai/methodology/intelligence-benchmarkingAccessed 6 September 2026
[8]Artificial Analysis. Announcing Harvey LAB-AA: evaluating AI agents on real-world legal work. 7 July 2026.https://artificialanalysis.ai/articles/harvey-lab-aaAccessed 6 September 2026
[9]Artificial Analysis, Inc.. Artificial Analysis Data Platform Terms and Conditions, version 1.1. revised 19 August 2026.https://artificialanalysiscdn.com/legal/ProDataPlatformTerms.pdfAccessed 6 September 2026
[10]Anthropic. System Card: Claude Opus 5. 24 July 2026, 198 pages.https://www.anthropic.com/claude-opus-5-system-cardAccessed 6 September 2026
[11]Anthropic. System Card: Claude Fable 5.1 & Claude Mythos 5.1. 1 September 2026, 212 pages.https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-cardAccessed 6 September 2026
[12]Anthropic. System Card: Claude Sonnet 5. 30 June 2026, 145 pages.https://www.anthropic.com/claude-sonnet-5-system-cardAccessed 6 September 2026
[13]Anthropic. Data retention practices for Covered Models. Anthropic Privacy Center, page carries a relative date only.https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-modelsAccessed 6 September 2026
[14]Anthropic. Covered Models. Anthropic Help Center, page carries a relative date only.https://support.claude.com/en/articles/15425695-covered-modelsAccessed 1 October 2026
[15]Anthropic. Data residency. Claude Platform Docs, undated.https://platform.claude.com/docs/en/manage-claude/data-residencyAccessed 6 September 2026
[16]Anthropic. API and data retention. Claude Platform Docs, undated.https://platform.claude.com/docs/en/manage-claude/api-and-data-retentionAccessed 28 September 2026
[17]Google Cloud. Abuse monitoring. Gemini Enterprise Agent Platform, updated 19 October 2026; as a Google Frontier Model only Gemini 4 Argon Cyber Permissive.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/abuse-monitoringAccessed 20 October 2026
[18]Google Cloud. Claude Opus 5 on Google Cloud. updated 2 September 2026.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5Accessed 6 September 2026
[19]Google Cloud. Claude Sonnet 5 on Google Cloud. updated 2 September 2026.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/sonnet-5Accessed 6 September 2026
[20]Google DeepMind. Gemini 3.8 Flash (Model evaluation: Approach, methodology & results). September 2026, 4 pages (PDF).https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdfAccessed 6 September 2026
[21]Microsoft. Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry. Microsoft Learn, 18 May 2026.https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacyAccessed 28 September 2026
[22]Microsoft. Deploy and use Grok models in Microsoft Foundry. Microsoft Learn, 26 August 2026.https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-grokAccessed 6 September 2026
[23]Microsoft. Region availability for Foundry Models sold by Azure. Microsoft Learn, page dated 3 September 2026; the table already lists the 22 September 2026 versions of gpt-6-sol and gpt-6-luna and the 29 September 2026 version of gpt-6.1-sol.https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availabilityAccessed 1 October 2026
[24]OpenAI. GPT-5.6 System Card. 9 July 2026, revised 19 August 2026, 82 pages (PDF).https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdfAccessed 6 September 2026
[25]OpenAI. GPT-6 Astra System Card. 3 September 2026, 117 pages (PDF).https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdfAccessed 6 September 2026
[26]SpaceXAI (DBA xAI LLC). Model Card: Grok 4.6. 12 August 2026, revised 17 August 2026, 42 pages (PDF).https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdfAccessed 6 September 2026
[27]Kimi Team (Moonshot AI). Kimi K3: Open Frontier Intelligence (Technical Report of Kimi K3). arXiv:2607.24653v2 [cs.CL], 7 August 2026, 47 pages.https://arxiv.org/abs/2607.24653Accessed 6 September 2026
[28]Meta AI Research. Introducing Muse Spark 1.3. 2 September 2026.https://research.meta.ai/blog/introducing-muse-spark-1-3Accessed 6 September 2026
[29]Meta. Muse Spark 1.3 (model page). undated.https://developer.meta.com/ai/models/muse-spark/Accessed 6 September 2026
[30]Meta. Meet Muse Spark 1.2 and Muse Code: a coding model and the agent built to run it. 5 August 2026.https://developer.meta.com/ai/resources/blog/build-with-muse-code/Accessed 6 September 2026
[31]Mistral AI. Mistral Medium 3.5 (model card, version v26.04). Modified MIT licence.https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04Accessed 6 September 2026
[32]Mistral AI. In-region inference, open models, and new European infrastructure for sovereign AI. 11 August 2026.https://mistral.ai/news/regional-inference-open-models-new-compute/Accessed 6 September 2026
[33]Mistral AI. Regional inference. Mistral Docs, undated.https://docs.mistral.ai/inference/regional-inferenceAccessed 6 September 2026
[34]Banar, N., Lotfi, E., Van Nooten, J., Kliocaite, M. & Daelemans, W.. bLLeQA: Benchmarking LLMs for Grounded Legal Question-Answering in French and Dutch. Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026), ACL, July 2026, pp. 34–59, DOI 10.18653/v1/2026.knowfm-1.4.https://aclanthology.org/2026.knowfm-1.4/Accessed 6 September 2026
[35]Ovcharov, V.. Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions. arXiv:2605.29738v2, 7 August 2026.https://arxiv.org/abs/2605.29738Accessed 6 September 2026
[36]European Commission, DG Translation and the EMT network. EU-MMLU (dataset). 17,203 rows, 16 language configurations including NL_NL, CC BY 4.0 licence.https://huggingface.co/datasets/EC-DGT-AI/EU-MMLUAccessed 6 September 2026
[37]Artificial Analysis. Model pages (Intelligence Index v4.3.2 per model). continuously updated model pages, values read on 1 October 2026 (Mistral Large 4: 20 October 2026), per model at the effort setting of the Vals run.https://artificialanalysis.ai/modelsAccessed 1 October 2026
[38]Google Cloud. Service Specific Terms. Gemini Enterprise Agent Platform / generative AI services: Google does not use customer data to train or fine-tune models without permission.https://cloud.google.com/terms/service-termsAccessed 28 September 2026
[39]Prudai B.V.. Privacy statement. Prudai Legal Center, chapter on subprocessors and processing locations.https://legal.prudai.com/privacyAccessed 6 September 2026
[40]Z.ai (zai-org). GLM-5.3 (model card on Hugging Face). 753 billion parameters, GLM-5.3 licence (commercial use allowed with restrictions), weights downloadable; accessed 22 September 2026.https://huggingface.co/zai-org/GLM-5.3Accessed 22 September 2026
[41]Z.ai (zai-org). GLM-5.3-Flash (model card on Hugging Face). 320 billion parameters, 18 billion active, MIT licence; accessed 22 September 2026.https://huggingface.co/zai-org/GLM-5.3-FlashAccessed 22 September 2026
[42]DeepSeek. DeepSeek-V4.1-Flash (model card on Hugging Face). 552 billion parameters, 8 to 16 billion active, MIT licence, released 10 September 2026; accessed 22 September 2026.https://huggingface.co/deepseek-ai/DeepSeek-V4.1-FlashAccessed 22 September 2026
[43]DeepSeek. DeepSeek-V4-Pro-0813 (model card on Hugging Face). 1.7 trillion parameters, MIT licence, 13 August 2026 release replacing the preview; accessed 22 September 2026.https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813Accessed 22 September 2026
[44]European open-weights host. Legal Quick Guide (provider documentation). undated documentation: public endpoints show region Global and the location can change without notice; dedicated endpoints fix the region contractually (DPA §6.1); no training on customer content; zero data retention is an account setting; accessed 22 September 2026.https://docs.tokenfactory.nebius.com/legal/legal-quick-guideAccessed 22 September 2026
[45]European open-weights host. Model catalogue of the provider. catalogue of 24 models, read through the API on 22 September 2026; including Kimi K3, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro 0813.https://tokenfactory.nebius.com/modelsAccessed 22 September 2026
[46]Prudai B.V.. Data Pro Statement, version 1.1. September 2026, article 5 (subprocessors) and the version history.https://legal.prudai.com/data-pro-statementAccessed 22 September 2026
[47]Google Cloud. Claude Opus 5.5 on Google Cloud. updated 25 September 2026; generally available since 22 September 2026, model availability and ML processing in the Europe multi-region.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5-5Accessed 28 September 2026
[48]Anthropic. Claude Opus 5.5. launch post, 22 September 2026: like earlier Opus models available with zero data retention.https://www.anthropic.com/claude-opus-5-5Accessed 28 September 2026
[49]Vals AI. Claude Opus 5.5 (model page). evaluation note of 22 September 2026: max effort, Claude Opus 5 and Claude Opus 4.8 as server-side fallbacks for refusals; on the Legal Research Bench 11 refusals without fallback, score unchanged.https://www.vals.ai/models/anthropic_claude-opus-5-5Accessed 28 September 2026
[50]Google. Gemini 4 Argon: our next era of frontier intelligence. launch post, 30 September 2026: first rolled out to participants in the Fairwind Program, then as soon as possible to paid API customers; introductory price 2 and 10 dollars per million tokens, then 4 and 20 dollars.https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/Accessed 1 October 2026
[51]Vals AI. Gemini 4 Argon (model page). evaluation note of 30 September 2026: high effort, temperature 1.0; Vals prices at 4 and 20 dollars per million tokens.https://www.vals.ai/models/google_gemini-4-argonAccessed 1 October 2026
[52]Google Cloud. Claude Sonnet 5.5 on Google Cloud. generally available since 28 September 2026, model availability and ML processing in the Europe multi-region.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/sonnet-5-5Accessed 1 October 2026
[53]Anthropic. Claude Sonnet 5.5. launch post, 28 September 2026: like Claude Opus 5.5 and Claude Sonnet 5 available with zero data retention.https://www.anthropic.com/claude-sonnet-5-5Accessed 1 October 2026
[54]Vals AI. Claude Sonnet 5.5 (model page). evaluation note of 28 September 2026: max effort, with Claude Sonnet 5 as a fallback on refusals; on the Legal Research Bench it stepped in 11 times (5.29 percent) and 3 refusals were scored as failures.https://www.vals.ai/models/anthropic_claude-sonnet-5-5Accessed 1 October 2026
[55]Google Cloud. Gemini 4 Argon. Gemini Enterprise Agent Platform, updated 20 October 2026: generally available since 13 October 2026, model availability and ML processing in the eu and us multi-regions, data residency supported.https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-argonAccessed 20 October 2026
[56]Google Cloud. Generative AI pricing. Gemini 4 Argon: through 30 November 2026 2 and 10 dollars per million tokens (global), from 1 December 4 and 20 dollars; outside global 10 percent higher.https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricingAccessed 20 October 2026
[57]Vals AI. Mistral Large 4 (model page). hyperparameters: provider Mistral, high effort, temperature 1.https://www.vals.ai/models/mistralai_mistral-large-4Accessed 20 October 2026
[58]Mistral AI. Introducing Mistral Large 4. launch post, 6 October 2026: public preview through the API, trained and served in Mistral’s own European data centres, a European deployment announced, open weights at the end of October.https://mistral.ai/news/mistral-large-4Accessed 20 October 2026
[59]Mistral AI SAS. Commercial Terms of Service. effective 25 September 2026; §4.2(d) and §4.3: for preview models Mistral may use customer data and outputs for training, and opt-out and zero data retention do not apply.https://legal.mistral.ai/terms/commercial-terms-of-service/Accessed 20 October 2026

More at Prudai

Full source list

[1] Vals AI. Legal Research Bench. benchmark page, version 1, updated 19 October 2026. https://www.vals.ai/benchmarks/legal_research. Accessed 20 October 2026.
[2] Vals AI. Harvey's Legal Agent Benchmark (HLAB leaderboard). leaderboard, version 1, archived, last updated 18 October 2026; since 19 October 2026 Vals shows version 1.1 by default, which covers part of this list. https://www.vals.ai/benchmarks/hlab-v1. Accessed 20 October 2026.
[3] Vals AI. Methodology. methodology page, undated. https://www.vals.ai/methodology. Accessed 6 September 2026.
[4] Grupen, N., Pereyra, G. & Pereyra, J. (Harvey AI). Open-Sourcing Harvey's Long Horizon Legal Agent Benchmark. Harvey AI, 6 May 2026. https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmark. Accessed 6 September 2026.
[5] Harvey AI. Harvey LAB: The Legal Agent Benchmark, v1.0. dataset and harness on GitHub, MIT licence. https://github.com/harveyai/harvey-labs. Accessed 6 September 2026.
[6] Artificial Analysis. Announcing the Artificial Analysis Intelligence Index v4.3. 7 September 2026; model pages show sub-version v4.3.2. https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3. Accessed 22 September 2026.
[7] Artificial Analysis. Artificial Analysis Intelligence Benchmarking Methodology. methodology page, undated. https://artificialanalysis.ai/methodology/intelligence-benchmarking. Accessed 6 September 2026.
[8] Artificial Analysis. Announcing Harvey LAB-AA: evaluating AI agents on real-world legal work. 7 July 2026. https://artificialanalysis.ai/articles/harvey-lab-aa. Accessed 6 September 2026.
[9] Artificial Analysis, Inc.. Artificial Analysis Data Platform Terms and Conditions, version 1.1. revised 19 August 2026. https://artificialanalysiscdn.com/legal/ProDataPlatformTerms.pdf. Accessed 6 September 2026.
[10] Anthropic. System Card: Claude Opus 5. 24 July 2026, 198 pages. https://www.anthropic.com/claude-opus-5-system-card. Accessed 6 September 2026.
[11] Anthropic. System Card: Claude Fable 5.1 & Claude Mythos 5.1. 1 September 2026, 212 pages. https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system-card. Accessed 6 September 2026.
[12] Anthropic. System Card: Claude Sonnet 5. 30 June 2026, 145 pages. https://www.anthropic.com/claude-sonnet-5-system-card. Accessed 6 September 2026.
[13] Anthropic. Data retention practices for Covered Models. Anthropic Privacy Center, page carries a relative date only. https://privacy.claude.com/en/articles/15425996-data-retention-practices-for-covered-models. Accessed 6 September 2026.
[14] Anthropic. Covered Models. Anthropic Help Center, page carries a relative date only. https://support.claude.com/en/articles/15425695-covered-models. Accessed 1 October 2026.
[15] Anthropic. Data residency. Claude Platform Docs, undated. https://platform.claude.com/docs/en/manage-claude/data-residency. Accessed 6 September 2026.
[16] Anthropic. API and data retention. Claude Platform Docs, undated. https://platform.claude.com/docs/en/manage-claude/api-and-data-retention. Accessed 28 September 2026.
[17] Google Cloud. Abuse monitoring. Gemini Enterprise Agent Platform, updated 19 October 2026; as a Google Frontier Model only Gemini 4 Argon Cyber Permissive. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/abuse-monitoring. Accessed 20 October 2026.
[18] Google Cloud. Claude Opus 5 on Google Cloud. updated 2 September 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5. Accessed 6 September 2026.
[19] Google Cloud. Claude Sonnet 5 on Google Cloud. updated 2 September 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/sonnet-5. Accessed 6 September 2026.
[20] Google DeepMind. Gemini 3.8 Flash (Model evaluation: Approach, methodology & results). September 2026, 4 pages (PDF). https://storage.googleapis.com/deepmind-media/gemini/gemini_3-8_flash_model_evaluation.pdf. Accessed 6 September 2026.
[21] Microsoft. Data, privacy, and security for Foundry Models sold by Azure in Microsoft Foundry. Microsoft Learn, 18 May 2026. https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/openai/data-privacy. Accessed 28 September 2026.
[22] Microsoft. Deploy and use Grok models in Microsoft Foundry. Microsoft Learn, 26 August 2026. https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-grok. Accessed 6 September 2026.
[23] Microsoft. Region availability for Foundry Models sold by Azure. Microsoft Learn, page dated 3 September 2026; the table already lists the 22 September 2026 versions of gpt-6-sol and gpt-6-luna and the 29 September 2026 version of gpt-6.1-sol. https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability. Accessed 1 October 2026.
[24] OpenAI. GPT-5.6 System Card. 9 July 2026, revised 19 August 2026, 82 pages (PDF). https://deploymentsafety.openai.com/gpt-5-6/gpt-5-6.pdf. Accessed 6 September 2026.
[25] OpenAI. GPT-6 Astra System Card. 3 September 2026, 117 pages (PDF). https://deploymentsafety.openai.com/gpt-6-astra/gpt-6-astra.pdf. Accessed 6 September 2026.
[26] SpaceXAI (DBA xAI LLC). Model Card: Grok 4.6. 12 August 2026, revised 17 August 2026, 42 pages (PDF). https://media.x.ai/v1/website/card-4p6-4cd2dc57.pdf. Accessed 6 September 2026.
[27] Kimi Team (Moonshot AI). Kimi K3: Open Frontier Intelligence (Technical Report of Kimi K3). arXiv:2607.24653v2 [cs.CL], 7 August 2026, 47 pages. https://arxiv.org/abs/2607.24653. Accessed 6 September 2026.
[28] Meta AI Research. Introducing Muse Spark 1.3. 2 September 2026. https://research.meta.ai/blog/introducing-muse-spark-1-3. Accessed 6 September 2026.
[29] Meta. Muse Spark 1.3 (model page). undated. https://developer.meta.com/ai/models/muse-spark/. Accessed 6 September 2026.
[30] Meta. Meet Muse Spark 1.2 and Muse Code: a coding model and the agent built to run it. 5 August 2026. https://developer.meta.com/ai/resources/blog/build-with-muse-code/. Accessed 6 September 2026.
[31] Mistral AI. Mistral Medium 3.5 (model card, version v26.04). Modified MIT licence. https://docs.mistral.ai/models/model-cards/mistral-medium-3-5-26-04. Accessed 6 September 2026.
[32] Mistral AI. In-region inference, open models, and new European infrastructure for sovereign AI. 11 August 2026. https://mistral.ai/news/regional-inference-open-models-new-compute/. Accessed 6 September 2026.
[33] Mistral AI. Regional inference. Mistral Docs, undated. https://docs.mistral.ai/inference/regional-inference. Accessed 6 September 2026.
[34] Banar, N., Lotfi, E., Van Nooten, J., Kliocaite, M. & Daelemans, W.. bLLeQA: Benchmarking LLMs for Grounded Legal Question-Answering in French and Dutch. Proceedings of the 4th Workshop on Towards Knowledgeable Foundation Models (KnowFM 2026), ACL, July 2026, pp. 34–59, DOI 10.18653/v1/2026.knowfm-1.4. https://aclanthology.org/2026.knowfm-1.4/. Accessed 6 September 2026.
[35] Ovcharov, V.. Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions. arXiv:2605.29738v2, 7 August 2026. https://arxiv.org/abs/2605.29738. Accessed 6 September 2026.
[36] European Commission, DG Translation and the EMT network. EU-MMLU (dataset). 17,203 rows, 16 language configurations including NL_NL, CC BY 4.0 licence. https://huggingface.co/datasets/EC-DGT-AI/EU-MMLU. Accessed 6 September 2026.
[37] Artificial Analysis. Model pages (Intelligence Index v4.3.2 per model). continuously updated model pages, values read on 1 October 2026 (Mistral Large 4: 20 October 2026), per model at the effort setting of the Vals run. https://artificialanalysis.ai/models. Accessed 1 October 2026.
[38] Google Cloud. Service Specific Terms. Gemini Enterprise Agent Platform / generative AI services: Google does not use customer data to train or fine-tune models without permission. https://cloud.google.com/terms/service-terms. Accessed 28 September 2026.
[39] Prudai B.V.. Privacy statement. Prudai Legal Center, chapter on subprocessors and processing locations. https://legal.prudai.com/privacy. Accessed 6 September 2026.
[40] Z.ai (zai-org). GLM-5.3 (model card on Hugging Face). 753 billion parameters, GLM-5.3 licence (commercial use allowed with restrictions), weights downloadable; accessed 22 September 2026. https://huggingface.co/zai-org/GLM-5.3. Accessed 22 September 2026.
[41] Z.ai (zai-org). GLM-5.3-Flash (model card on Hugging Face). 320 billion parameters, 18 billion active, MIT licence; accessed 22 September 2026. https://huggingface.co/zai-org/GLM-5.3-Flash. Accessed 22 September 2026.
[42] DeepSeek. DeepSeek-V4.1-Flash (model card on Hugging Face). 552 billion parameters, 8 to 16 billion active, MIT licence, released 10 September 2026; accessed 22 September 2026. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash. Accessed 22 September 2026.
[43] DeepSeek. DeepSeek-V4-Pro-0813 (model card on Hugging Face). 1.7 trillion parameters, MIT licence, 13 August 2026 release replacing the preview; accessed 22 September 2026. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813. Accessed 22 September 2026.
[44] European open-weights host. Legal Quick Guide (provider documentation). undated documentation: public endpoints show region Global and the location can change without notice; dedicated endpoints fix the region contractually (DPA §6.1); no training on customer content; zero data retention is an account setting; accessed 22 September 2026. https://docs.tokenfactory.nebius.com/legal/legal-quick-guide. Accessed 22 September 2026.
[45] European open-weights host. Model catalogue of the provider. catalogue of 24 models, read through the API on 22 September 2026; including Kimi K3, GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash and DeepSeek V4 Pro 0813. https://tokenfactory.nebius.com/models. Accessed 22 September 2026.
[46] Prudai B.V.. Data Pro Statement, version 1.1. September 2026, article 5 (subprocessors) and the version history. https://legal.prudai.com/data-pro-statement. Accessed 22 September 2026.
[47] Google Cloud. Claude Opus 5.5 on Google Cloud. updated 25 September 2026; generally available since 22 September 2026, model availability and ML processing in the Europe multi-region. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5-5. Accessed 28 September 2026.
[48] Anthropic. Claude Opus 5.5. launch post, 22 September 2026: like earlier Opus models available with zero data retention. https://www.anthropic.com/claude-opus-5-5. Accessed 28 September 2026.
[49] Vals AI. Claude Opus 5.5 (model page). evaluation note of 22 September 2026: max effort, Claude Opus 5 and Claude Opus 4.8 as server-side fallbacks for refusals; on the Legal Research Bench 11 refusals without fallback, score unchanged. https://www.vals.ai/models/anthropic_claude-opus-5-5. Accessed 28 September 2026.
[50] Google. Gemini 4 Argon: our next era of frontier intelligence. launch post, 30 September 2026: first rolled out to participants in the Fairwind Program, then as soon as possible to paid API customers; introductory price 2 and 10 dollars per million tokens, then 4 and 20 dollars. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/. Accessed 1 October 2026.
[51] Vals AI. Gemini 4 Argon (model page). evaluation note of 30 September 2026: high effort, temperature 1.0; Vals prices at 4 and 20 dollars per million tokens. https://www.vals.ai/models/google_gemini-4-argon. Accessed 1 October 2026.
[52] Google Cloud. Claude Sonnet 5.5 on Google Cloud. generally available since 28 September 2026, model availability and ML processing in the Europe multi-region. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/sonnet-5-5. Accessed 1 October 2026.
[53] Anthropic. Claude Sonnet 5.5. launch post, 28 September 2026: like Claude Opus 5.5 and Claude Sonnet 5 available with zero data retention. https://www.anthropic.com/claude-sonnet-5-5. Accessed 1 October 2026.
[54] Vals AI. Claude Sonnet 5.5 (model page). evaluation note of 28 September 2026: max effort, with Claude Sonnet 5 as a fallback on refusals; on the Legal Research Bench it stepped in 11 times (5.29 percent) and 3 refusals were scored as failures. https://www.vals.ai/models/anthropic_claude-sonnet-5-5. Accessed 1 October 2026.
[55] Google Cloud. Gemini 4 Argon. Gemini Enterprise Agent Platform, updated 20 October 2026: generally available since 13 October 2026, model availability and ML processing in the eu and us multi-regions, data residency supported. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/4-argon. Accessed 20 October 2026.
[56] Google Cloud. Generative AI pricing. Gemini 4 Argon: through 30 November 2026 2 and 10 dollars per million tokens (global), from 1 December 4 and 20 dollars; outside global 10 percent higher. https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing. Accessed 20 October 2026.
[57] Vals AI. Mistral Large 4 (model page). hyperparameters: provider Mistral, high effort, temperature 1. https://www.vals.ai/models/mistralai_mistral-large-4. Accessed 20 October 2026.
[58] Mistral AI. Introducing Mistral Large 4. launch post, 6 October 2026: public preview through the API, trained and served in Mistral’s own European data centres, a European deployment announced, open weights at the end of October. https://mistral.ai/news/mistral-large-4. Accessed 20 October 2026.
[59] Mistral AI SAS. Commercial Terms of Service. effective 25 September 2026; §4.2(d) and §4.3: for preview models Mistral may use customer data and outputs for training, and opt-out and zero data retention do not apply. https://legal.mistral.ai/terms/commercial-terms-of-service/. Accessed 20 October 2026.

Claude and Anthropic, GPT and OpenAI, Gemini and Google, Grok and xAI, Kimi and Moonshot AI, Muse and Meta, Mistral, DeepSeek, GLM and Z.ai, Harvey, Vals AI and Artificial Analysis are trademarks of their respective owners. They are named for identification only and this implies no partnership, sponsorship or endorsement.

The logos shown are the property of their respective owners and appear here purely to identify the provider or the route. The icon set comes from Simple Icons (CC0).

Beau Jonkhout, Chief Technology Officer at Prudai · Last updated 20 October 2026 · PrudentBench v1.1

Which AI model is suitable for Dutch legal research?

1. The scorecard

PrudentScore v1.1

2. Why yes, why no: model by model

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

Strong

Weak

3. Compliance matrix

Why the gate closes

4. How we choose models

01Quality on legal research

02Legal agent tasks

03Cost per unit of work

04Speed

05The EU route

06Reliability is architecture

The six building blocks of a choice

5. Methodology and PrudentScore

The gate

PrudentScore v1.1

Results for v1.1, computed from this formula

What we do ourselves

What we do not measure

6. What changed

7. Frequently asked questions

What is the best AI model for legal advice?

Is ChatGPT GDPR-proof for law firms?

May I enter client data into an AI model?

Claude or ChatGPT for legal research?

Why does AI invent case law, and which model does it least?

Which models process data in the EU?

Are open-weights models automatically compliant?

Does the provider train on my case files?

What does the AI Act mean for my firm?

How often is PrudentBench updated?

Does Prudai measure itself, or only summarise?

8. Sources

Quality on legal research

Legal agent tasks

Cost per unit of work

Speed

The EU route

Reliability is architecture