prinzbench is a private benchmark that ranks LLMs based on their ability to conduct legal research and analysis ("legal reasoning") and locate obscure publicly available information online ("needle-in-the-haystack search").

The current leaderboard is set forth immediately below, followed by notes on methodology, grading, model access and specific models' performance.

IMPORTANT UPDATE: This benchmark was created in January 2026, with the goal of testing frontier AI models' ability to conduct legal research. In June 2026, gpt-5.6-Sol-Pro achieved a score of 91/99 on prinzbench, effectively saturating it. My benchmark lasted 6 months! I extend my heartfelt congratulations to OpenAI - gg!

Last updated on July 18, 2026

* denotes a model accessed in early testing, before public release.

Why this benchmark was created:

1. There has recently been much interest in benchmarks that measure LLMs' ability to perform economically valuable work. The subject of prinzbench - legal research - is economically valuable work.
2. Many existing LLM benchmarks focus on subjects like math and coding, and there has recently been a sharp increase in LLMs' performance on these benchmarks. However, some skeptics have declared that existing LLMs are great at math and coding only because the frontier labs overfit the models for these specific tasks, "hill-climbing the benchmarks" even as AI models' ability to help with other real-world tasks has remained limited. prinzbench proves these skeptics wrong by showing that certain existing LLMs possess surprisingly strong ability to correctly research and analyze statutes, regulations and regulatory guidance in a particular niche area of U.S. law.²

prinzbench is intended to test AI models' ability to: (i) find legal authorities or factual information that are relevant to answering a legal research question; and (ii) properly analyze available legal authorities to reach the correct conclusion. To that end, prinzbench comprises:

• 25 Legal Research questions, each of which presents the LLM with a fact pattern that must be correctly analyzed based on applicable U.S. laws
• 8 Search questions, each of which asks the LLM to find an obscure piece of information that is publicly available online

All 33 questions were drafted by the author, and have never been - nor will they ever be - shared with any other person.

The Legal Research questions all relate to a particular niche area of U.S. law in which the author is an expert. To answer these questions correctly, the LLM is required to locate the relevant statutes, regulations and/or regulatory guidance - all of which are publicly available online. The answer to each Legal Research question can be found by examining only a few applicable legal authorities; there is never any need for the LLM to trawl through hundreds of cases and juxtapose their holdings in order to arrive at the correct response. Certain Legal Research questions also require the LLM to locate obscure publicly available factual information to reach the correct conclusion.

The Search questions were added to prinzbench to strengthen the benchmark's ability to test LLMs for "needle-in-the-haystack" search capabilities. These questions are not necessarily related to the practice of law.

All questions included in prinzbench are considered by the author to be very difficult: generally of the kind that, in the author's judgment, an average junior associate practicing in the relevant area of the law would likely not be able answer correctly if provided no research direction beyond the prompt provided to the LLM. In compiling prinzbench, the author endeavored to include therein only truly challenging questions, and each candidate question was tested on an ad-hoc basis with various LLMs before a decision was made regarding its inclusion in the benchmark. Well over 100 candidate questions were rejected through this process, as the LLMs - frustratingly for the author - had found them "too easy".

To arrive at each model's prinzbench (full) score, the model was asked each of the 33 questions exactly three (3) times. Each response was graded at pass@1. The maximum prinzbench (full) score is 99 points (i.e., 33 questions * 3).

In addition to each model's prinzbench (full) score, the leaderboard also presents each model's Legal Research sub-score (maximum 75 points) and Search sub-score (maximum 24 points), for informational purposes.

Each LLM's response to each question was personally graded by the author, unaided by AI. The author graded each output with full knowledge of which model had produced the output (i.e., grading was not "blind").

Some technical details regarding grading:

• A response was judged to be correct if the correct legal conclusion was present anywhere within the model's response. For example, suppose the correct answer to a particular question is: "You must aggregate the impact of not fewer than 1,000 different transactions of the same type in order to violate a law; because it's unlikely that there would ever be 1,000 such transactions within the specified time frame, it is almost certain that the law was not violated." If this exact reasoning was included in the body of the model's response, but the model's concluding paragraph simply said that the law was not violated, the response was marked as correct.
• LLMs have a tendency to include extraneous information in their responses, which is not directly relevant to answering the question. In grading LLMs' responses, all such extraneous information was ignored - regardless of whether it was factually accurate or legally correct. In other words, the models were graded on their ability to answer the specific question, not on the accuracy of the entirety of the response.

The following models will no longer be tested on prinzbench:

• After gpt-5.6-Sol-Pro scored 91/99 on prinzbench in June 2026, OpenAI's Pro models are no longer being tested on prinzbench.

In June 2026, gpt-5.6-Sol-Pro saturated prinzbench, with an overall score of 91/99. The model achieved a perfect Search sub-score (24/24).

The author notes that prinzbench contains two questions that no model tested to date has ever been able to solve. One of these questions involves an extremely thorough 50-state search, and may therefore be too difficult for modern models to answer without running in /goal mode (or similar) in an agentic harness. The other question requires the model to identify a total of four required regulatory approvals (one easy, two very difficult and one extremely difficult); gpt-5.6-Sol-Pro was able to find the one easy and two very difficult approvals, but could not find the extremely difficult approval. Putting these two questions (which are worth a total of 6 points) aside, gpt-5.6-Sol-Pro provided correct responses to 91 out of 93 prinzbench questions - an extremely impressive performance.

Other GPT-5.6 models will be tested on prinzbench separately; please stay tuned.

While the scores achieved by OpenAI's Pro models on prinzbench are impressive, the author cautions against directly comparing them to scores achieved by the other models tested to date, since the Pro models are extremely heavy reasoning models that use parallel test-time compute.³

At the benchmark's creation in January 2026, gpt-5.2-Thinking achieved the highest score on prinzbench by a large margin (52/99, as compared to the next-best model's 36/99). The model also achieved by far the best Legal Research and Search sub-scores (41/75 and 11/24, respectively).

In February 2026, gpt-5.3-Codex-high matched the high score previously achieved by gpt-5.2-Thinking*. While the scores and sub-scores achieved by the two models were, in the end, identical, the author found them to have different strengths and weaknesses. For example, gpt-5.3-Codex-high scored 3/3 on one particular question that had never been solved correctly by any other model, including gpt-5.2-Thinking*.

In March 2026, gpt-5.4-Thinking achieved better scores than gpt-5.2-Thinking across the board. Specifically, gpt-5.4-Thinking (Extended) scored 6 points higher on prinzbench than gpt-5.2-Thinking (Extended) (58/99 vs. 52/99), while gpt-5.4 (xhigh) achieved the then-best result for a non-Pro model (69/99). In addition, gpt-5.4-Pro (Extended) achieved an overall score of 79/99 on prinzbench - which was the best performance by any model on the benchmark through that date. In particular, the gpt-5.4 models performed remarkably well on the Search component of prinzbench (20/24 for gpt-5.4-Pro (Extended) and 19/24 for each of gpt-5.4-Thinking (Heavy) and gpt-5.4 (xhigh)).

In April 2026, In April 2026, gpt-5.5-Pro (Extended) set a new record on prinzbench, scoring 82/99 overall, having performed significantly better than gpt-5.4-Pro (Extended) in Legal Research (63/75 vs. 59/75), but slightly worse in Search (19/24 vs. 20/24). In addition, gpt-5.5-Thinking (Heavy), scored 74/99 on prinzbench - the best result by a non-Pro model.

gpt-5.6-Sol-Pro was accessed from the author's personal ChatGPT Pro account in June 2026. gpt-5.2-Thinking and gpt-5.4-Thinking ("Extended Thinking") were accessed from the author's personal ChatGPT Plus account in January 2026 and March 2026, respectively. The gpt-5.5 models and the gpt-5.4 models (in each case, using the "Extended" reasoning effort for the Pro models and the "Heavy" reasoning effort for the Thinking models) were accessed from the author's personal ChatGPT Pro account in April 2026 and March 2026, respectively. Each chat was a temporary chat in order to disable ChatGPT's Memory and custom instructions features. gpt-5.3-Codex-high and gpt-5.4-Thinking-xhigh were accessed by the author through the terminal, with his personal ChatGPT Plus/Pro accounts connected, in February 2026 and March 2026, respectively.

The author had early access to gpt-5.6-Sol-Pro, gpt-5.5-Pro and gpt-5.5-Thinking; see footnote 1 for more details.

At the benchmark's creation in January 2026, gemini-3-flash and gemini-3-pro achieved the second and third place, respectively, on prinzbench, bested only by gpt-5.2-Thinking. The two models' scores were nearly identical (36/99 and 35/99, respectively). The models exhibited especially strong Legal Research capabilities, bested only by gpt-5.2-Thinking, and far stronger than those of the tested Anthropic, xAI and Moonshot models.

gemini-3.1-pro, benchmarked in February 2026, nearly matched OpenAI's models (gpt-5.2-Thinking and gpt-5.3-Codex-high) in performance (50/99 vs. 52/99 for each of the OpenAI models). The author views the prinzbench scores achieved by these three models as roughly identical.

gemini-3-flash and gemini-3-pro were both accessed, in January 2026, via AI Studio, with the following settings: Temperature: 1; Thinking Level: High; Grounding with Google Search: On. gemini-3-pro was accessed via AI Studio using the same settings in February 2026.

In June-July 2026, fable-5 (max) achieved by far the best score by any non-OpenAI model on prinzbench to date, scoring 56/99. The model achieved an excellent Legal Research score (49/75), but scored poorly on Search (7/24).

The model's performance on Legal Research (49/75) tied with gpt-5.4-Thinking (heavy), but was worse than gpt-5.5-Thinking (heavy) (56/75). It is important to remember that the prinzbench Legal Research score depends in part on a model's Search capabilities, because answering a Legal Research question correctly requires the model to: (i) find the relevant legal authorities online (i.e., perform a web search), and (ii) correctly interpret these legal authorities. The author believes that, if fable-5 were equipped with better web search capabilities, its Legal Research score would improve beyond 49/75 - although it is not clear whether, in such circumstances, it would perform below, at, or above the gpt-5.5-Thinking (heavy) level. Disregarding fable-5's poor web search capabilities, the author is inclined to generally view fable-5 (max) and gpt-5.5-Thinking (heavy) as models of approximately the same class (with a slight preference towards the latter).

Between January and April 2026, Anthropic models (opus-4.5 through opus-4.7 and sonnet-4.5 through sonnet-4.6) achieved poor scores on prinzbench, with opus-4.7 scoring the highest (25/99). These models performed particularly poorly on Search, with opus-4.7 scoring the highest (4/24).

In June 2026, opus-4.8 (max) achieved the best score by an Anthropic model on prinzbench to date, scoring 42/99. The author speculates that much of the gain in performance between opus-4.7 and opus-4.8 (25/99 vs. 42/99) was due to the Max reasoning setting that was not available for the former model, but became available for the latter.

opus-4.5 and sonnet-4.5 were both accessed from the author's personal Claude Pro account, with "Extended Thinking" turned on, in January 2026. opus-4.6 and sonnet-4.6 were accessed from the author's personal Claude Pro account, with "Extended Thinking" turned on, in February 2026. opus-4.7 was accessed from the author's personal Claude Pro account, with "Extended Thinking" turned on, in April 2026. opus-4.8 was accessed form the author's personal Claude Pro account, with "Max" reasoning effort turned on, in late May/early June 2026. fable-5 was accessed form the author's personal Claude Pro account, with "Max" reasoning effort turned on, in June 2026 and (following the model's re-release) in early July 2026. Each chat was a temporary chat in order to disable Claude's Memory and custom instructions features.

At the benchmark's creation in January 2026, grok-4.1-Thinking and grok-4 achieved unspectacular scores on prinzbench of 25/99 and 23/99 respectively. The models performed as well as the Google models on the Search sub-score (6/24 and 7/24, respectively - which were the same results as those achieved by gemini-3-flash and gemini-3-pro), but significantly underperformed the Google models on the Legal Research sub-score.

In February 2026, grok-4.20 achieved the fourth-highest score on prinzbench (43/99). This model appears to take the "best of" the results independently achieved by four different agents, which makes it structurally different from all other models benchmarked through February 2026. This makes it difficult to directly compare the underlying model's performance to the respective results of all other benchmarked models.⁴

grok-4.1-Thinking and grok-4 were both accessed from the author's personal SuperGrok account in January 2026, and grok.4.20 was accessed in the same fashion in February 2026. Each chat was a temporary chat in order to disable Memory and custom instructions features.

In July 2026, kimi-K3 achieved the best score by an open-source model on prinzbench to date, scoring 47/99. kimi-K3 was the first non-OpenAI model to get 10 or more questions correct on the Search sub-component of the benchmark (11/24).

The author observes that kimi-K3's responses were very inconsistent in quality: at times, the model struggled with fairly easy prinzbench questions that most other modern models are able to nearly always answer correctly; but, on the other hand, some of kimi-K3's responses to much harder questions were very impressive.

Subjectively, the author views this model as roughly similar to gemini-3.1-pro (released in February 2026) in legal research capabilities (albeit slightly better at search, perhaps slightly worse at logical reasoning, and somewhat more inconsistent).

On January 28, 2026, kimi-K2.5-Thinking achieved an impressive score of 35/99 on prinzbench, tying the second-place scores previously achieved by the Gemini 3 models. At the benchmark's creation in January 2026, kimi-K2-Thinking achieved an unspectacular score of 22/99 on prinzbench. The model's performance was similar to the performance of the xAI models, although it slightly lagged them in the Search sub-score (4/24 vs. 7/24 and 6/24 for grok-4 and grok-4.1-Thinking, respectively).

kimi-K3 was accessed via kimi.com in July 2026 via kimi.com (logged in), and was set at Max reasoning effort. kimi-K2.5-Thinking and kimi-K2-Thinking were accessed via kimi.com (logged in) in January 2026.

In January 2026, qwen3-max achieved an unspectacular score of 25/99 on prinzbench, tying with grok-4.1-Thinking. The model performed significantly worse than kimi-K2.5-Thinking, which was released by Moonshot AI on the same day.

qwen3-max was accessed via chat.qwen.ai (logged in) in January 2026.

In April 2026, meta-muse-spark achieved an unspectacular score of 31/99 on prinzbench, several points lower than gemini-3-pro and significantly lower than those achieved by various models subsequently released by OpenAI and Google. The model's Search score (7/24) was comparable to Search scores achieved by various models released in late 2025/early 2026 (such as gemini-3-pro and kimi-k2-thinking).

meta-muse-spark was accessed via meta.ai (logged in) in April 2026.

In April/May 2026, deepseek-v4-pro achieved a low score of 23/99 on prinzbench, in line with old models like grok-4 and kimi-k2-thinking (each released in June/July 2025).

deepseek-v4-pro was accessed via chat.deepseek.com (logged in) in April/May 2026.

In June 2026, glm-5.2 achieved a low score of 30/99 on prinzbench, beating the likes of deepseek-v4-pro, but significantly trailing then-current frontier Western models (and even trailing the likes of gemini-3-pro, released 8 months earlier).

glm-5.2 was accessed via chat.z.ai. All prompts were run at Deep Think Max settings, with Advanced Search on.