Live Nov 18 · Cohort 3Nov 18, 19:00 MSK — free hands-on AI agents live workshop · Live Cohort 3 starts Nov 23Join Nov 18 Live →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableOpenAI Blog→ original

OpenAI Releases Legal AI: 95% Fewer Fabricated Facts Than Claude Opus 5.5

OpenAI has released Astra for Law, a version of GPT-6.2 Astra with legal search and built-in instructions. When a lawyer asks it to draft a document, the model is less likely to cite a case that doesn't exist. Across 120 test tasks, it produced 0.03 fabricated material facts per task versus 0.59 for Claude Opus 5.5, a 95% reduction. It earned a full score on 20.8% of tasks. Teams where a bad citation is costly should request access and test the model on their own tasks.

AI-processed from OpenAI Blog; edited by Hamidun News
OpenAI Releases Legal AI: 95% Fewer Fabricated Facts Than Claude Opus 5.5
Source: OpenAI Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceOpenAI Blog ↗
Editor-in-Chief
Cost~64K tokens per task versus 139K for Muse Wonder Full
Key metric #2The search index grew by 34% and covers 99.9% of US precedents
AccessChatGPT Enterprise customers through the Trusted Access Program

Imagine a lawyer asks an AI to prepare a court filing, and the model confidently cites a case that doesn't exist. One such fabricated citation can cost a lawyer a client's trust and their standing before the court. OpenAI has released Astra for Law, a legal version of its flagship GPT-6.

2 Astra, with a primary focus on preventing exactly this kind of fabrication. In an independent Artificial Analysis test of 120 tasks across 24 areas of law, it averaged 0.03 material fabrications per task, or roughly one per 33 tasks.

Claude Opus 5.5, the best non-OpenAI model, scored 0.59, nearly one for every two tasks.

That is a difference of roughly 20x, or 95%. Any ChatGPT Enterprise customer can now request access.

What the product is

Astra for Law is not a new model but a "configuration" of GPT-6.2 Astra. The base model is connected to an expanded legal search index and given instructions on how to research a question, weigh sources and write an answer. When a new base model ships, the legal version inherits its capabilities, while OpenAI continues to refine the settings based on feedback from lawyers.

On Harvey's own test (Legal Agent Benchmark), GPT-6 Astra fully met all requirements on 4.6% of tasks, GPT-6.2 Astra on 14.6%, and the legal configuration on 16.2%.

Since September, when the product was shown to a small group of firms (including Cooley, Ropes & Gray and Sullivan & Cromwell), the index has grown by 34%: more than 80 million links were added, including UK, Australian and Canadian law. It now holds more than 310 million URLs and covers more than 99.9% of published US case law.

What the tests showed

In LAB-AA, Artificial Analysis's version of the Harvey test, models tackle tasks ranging from deposition outlines to lease reviews. A full score is awarded only if every requirement is met and there are no material fabrications. Astra for Law got a full score on 20.8% of tasks. The standard GPT-6.2 Astra scored 15.0% (a gap of almost 40%), and the best third-party model, Muse Wonder Full, scored 13.1% (a gap of about 59%).

Cost matters for budgets, too. Astra for Law uses about 64 thousand output tokens per task, while Muse Wonder Full uses 139 thousand, which is 54% fewer. With usage-based pricing, that can decide whether a legal department stays within budget.

Astra for Law makes 0.06 minor fabrications per task, versus 0.51 for Claude Opus 5.5. Tasks that received a full score contained no material fabrications at all.

In drafting adversarial documents (LAB: USCL, created by Harvey and Grubel), the lead is narrower. When the model finds the relevant cases itself, Astra for Law scored 1,844 points versus 1,817 for Opus 5.5. When precedents are supplied in advance, the results are nearly identical: 1,883 and 1,882. Harvey's evaluators preferred the new version's drafts to GPT-6 Astra's in almost 90% of cases. On Vals AI's Legal Research Bench, the share of questions with a full score rose from 46% to 53.8% (112 of 208) in less than two months, and counting partial credit, the model scored 90.9%.

How to use it and what to keep in mind

OpenAI describes use cases from its own legal team. A competition lawyer asks the model to argue the opposite interpretation of a law, drawing on legislative history and court decisions. A product lawyer builds a map of the rules for an unfamiliar industry, so the first conversation with outside counsel starts with real questions. A legal program manager turns regulatory procedures into a workflow: when to request data, when to close a case, when to hand it to a lawyer.

Access goes to ChatGPT Enterprise customers through the Trusted Access Program, including for use via API.

There are caveats. OpenAI itself published the results, and some of the tests (for example, Harvey's LAB) were run by its partners. A full score of 20.8% means that roughly four tasks out of five still were not solved flawlessly. In document drafting, the advantage over Opus 5.5 is small.

The practical takeaway for legal-tech teams and legal departments: request access, but test the model on your own typical tasks. Check every citation against the primary source, and track token costs as well as quality. Where a fabricated quote is a reputational risk, a low fabrication rate matters more than polished prose.

Story

Key Questions Answered

What the product is?

Astra for Law is not a new model but a "configuration" of GPT-6.2 Astra. The base model is connected to an expanded legal search index and given instructions on how to research a question, weigh sources and write an answer. When a new base model ships, the legal version inherits its capabilities, while OpenAI continues to refine the settings based on feedback from lawyers.

What the tests showed?

In LAB-AA, Artificial Analysis's version of the Harvey test, models tackle tasks ranging from deposition outlines to lease reviews. A full score is awarded only if every requirement is met and there are no material fabrications. Astra for Law got a full score on 20.8% of tasks. The standard GPT-6.2 Astra scored 15.0% (a gap of almost 40%), and the best third-party model, Muse Wonder Full, scored 13.1% (a gap of about 59%). Cost matters for budgets, too.

How to use it and what to keep in mind?

OpenAI describes use cases from its own legal team. A competition lawyer asks the model to argue the opposite interpretation of a law, drawing on legislative history and court decisions. A product lawyer builds a map of the rules for an unfamiliar industry, so the first conversation with outside counsel starts with real questions. A legal program manager turns regulatory procedures into a workflow: when to request data, when to close a case, when to hand it to a lawyer.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…