On November 17, 2026, OpenAI released Astra for Law 6.2 and disclosed same-day evaluation results on the Harvey LAB-AA v1.1 benchmark: 120 private legal tasks covering 24 practice areas, with a full-task pass rate of 20.8%, ranking first among all evaluated models. This is the highest public score to date on the same benchmark; the general-purpose GPT-6.2 Astra on the same base scored 15.0% in the same evaluation, a gap of 5.8 percentage points. ChatGPT Enterprise customers can currently apply for access through OpenAI's trusted access program.
What it actually is, and how it outperformed its own "sibling"
Astra for Law is not a completely new model but a legal-specific configuration of GPT-6.2 Astra—understandable as system prompts, tool-calling chains, and output-format constraints for legal scenarios layered on top of the general-purpose base. This architectural difference is directly reflected in the evaluation numbers: on the same base, the legal-specific configuration is about 39% higher than the general-purpose configuration. This shows that the leverage of prompt engineering and task-specific alignment in legal scenarios is far greater than imagined.
The evaluation design of Harvey LAB-AA v1.1 is quite stringent. The task set was designed by legal AI company Harvey and covers real lawyers' daily work, from developing ownership-change strategies in M&A to commercial lease review and drafting indictment outlines; scoring uses a binary standard—all rubric criteria for each task must pass for it to count as a "full pass," and partial credit does not count. A 20.8% full-pass rate means that out of 100 tasks, only about 21 can get three AI reviewers to unanimously agree that they "fully meet the standard." That number may not sound high, but in the stringent context of legal work, it is already the ceiling for currently public models.
Beyond pass rate, the more notable metric in this evaluation is hallucination rate. According to Crypto Briefing, citing evaluation data, Astra for Law 6.2 had only 0.03 material hallucinations per task, the lowest among evaluated models, which works out to about three material errors per 100 tasks. The specific scope of this check is: comparing model outputs with the provided case documents and checking for statements unsupported by the documents—it does not verify whether case-law citations are accurate, but for legal representation work centered on "drafting documents based on files," this already covers the most critical error type.
Gains and losses for stakeholders
For legal practitioners, hallucination rate is the real threshold determining whether a tool can enter the workflow, not the pass-rate ranking. A lawyer's core professional risk is submitting incorrect facts to a client or court. A tool that can stably complete document drafting with low hallucinations can take on a large amount of first-draft work; a model with a high hallucination rate requires full review every time it is used, and the actual time saved is limited. From this perspective, the 0.03 material-hallucination rate is the most substantive data point in Astra for Law's release.
For enterprise users that have already deployed ChatGPT Enterprise, the access path is not immediately available but requires an active application. There are two reasonable explanations for this design: first, OpenAI is controlling capacity to avoid high-risk outputs in legal scenarios triggering compliance incidents at scale; second, it uses the application process to screen use cases and accumulate traceable real-world feedback. At the same time, legal technology partners such as Harvey and Legora will connect via API, meaning users of these platforms will become another major channel in the future.
For the competitive landscape, this release creates a structural pressure point: the general-purpose base is already approaching the usable threshold in legal scenarios, which will compress the differentiation room for vertical legal AI companies that sell specifically to law firms. When the general-purpose GPT-6.2 Astra itself also reaches a 15% full-pass rate, the selling point of "designed specifically for law" needs to be supported by more concrete, actionable advantages rather than broad accuracy promises.
How much can one benchmark explain?
Harvey LAB-AA is not the only legal AI evaluation framework, and putting results from several benchmarks together makes the picture much more complex. According to reports, on another HLAB benchmark, GPT-6.2 Astra ranked only 11th; on a legal research benchmark it ranked 30th. In Vals AI's legal research benchmark, Astra for Law 6.2 achieved a 53.8% full-score rate on November 16, 2026, estimated to be about 8 percentage points higher than the baseline two months earlier, but it did not top the overall ranking on that benchmark.
This cross-benchmark divergence is not surprising, but it needs to be treated cautiously. Harvey LAB-AA tests task-based agentic capability—given a set of materials, complete a specific legal document—while legal research benchmarks focus more on retrieving and citing case law and arguing legal application. Both capabilities are equally important for legal work, but training optimization can point in different directions. Astra for Law's first place on Harvey LAB-AA shows that it has made clear task-specific alignment for document-drafting tasks, but this is not equivalent to comprehensive leadership in all-around legal AI capabilities.
Harvey LAB-AA's task set was designed by Harvey, and Harvey itself is a commercial legal AI company and one of OpenAI's legal-domain partners. Independent execution of the evaluation was carried out by Artificial Analysis, but authority over task design and scoring criteria belongs to Harvey, which logically creates a structural issue of stakeholder involvement in benchmark design. This does not mean the data is fabricated, but when interpreting "first place," it needs to be understood within the limited context of "a task set designed by Harvey."
What to watch next
From the infrastructure layer, GPT-6.2 Astra officially launched on Amazon Bedrock the same day, November 17, with API pricing at $8 per million input tokens and $40 per million output tokens, 20% lower than the previous-generation GPT-6 Astra. This means enterprise technology teams can now call the base model through standard cloud services without relying entirely on OpenAI's direct interface. Astra for Law's enterprise access and the underlying API availability landed simultaneously, which is a complete package from an engineering deployment perspective.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接