OpenAI updated Astra for Law on November 17. The legal product now runs on GPT-6.2 Astra, and every ChatGPT Enterprise customer can request access, with API access through a separate request form. On Artificial Analysis’s run of Harvey’s Legal Agent Benchmark (LAB-AA) it ranks first of 45 listed model configurations, earning full credit on 20.8% of assignments against 15.0% for plain GPT-6.2 Astra. OpenAI’s post gives no price. OpenAI post · LAB-AA leaderboard
The LAB-AA and Vals figures below are from those organizations’ leaderboards as checked on November 17. The index size, output-token counts, LAB: USCL scores and Harvey’s in-house LAB results come from OpenAI’s post, and we found no outside confirmation of them. S5 Labs has not run or tested Astra for Law.
What a model configuration is
Astra for Law launched on September 17 on GPT-6 Astra. LawNext reported then that Jason Boehmig, who leads OpenAI’s legal product development, stressed it was a configuration of an existing model, not a new one: domain-specific instructions, response-length settings and legal tooling. The November post describes the same structure on GPT-6.2 Astra, pairing the model with a larger legal search index and instructions for researching a question, weighing authorities and writing the answer. LawNext · OpenAI post
In September the index held more than 230 million URLs, drew on CourtListener and covered U.S. sources only. OpenAI now says it spans over 310 million URLs, 34% more, with UK, Australian and Canadian material among the additions, and that it covers more than 99.9% of published U.S. precedential case law, a figure mixed-news also reported in September. We found no third-party check of that coverage figure. LawNext · mixed-news · OpenAI post
Because the product is a general model plus retrieval and instructions, a benchmark that lists both the base model and Astra for Law shows what the configuration adds:
| Benchmark | GPT-6 Astra | GPT-6.2 Astra | Astra for Law | Reported by |
|---|---|---|---|---|
| LAB-AA, full credit (Max effort) | 8.6% | 15.0% | 20.8% | Artificial Analysis |
| Harvey’s own LAB, every requirement met | 4.6% | 14.6% | 16.2% | OpenAI |
| Vals Legal Research Bench, full credit | 39.42% | 42.79% | 53.85% | Vals AI |
LAB-AA leaderboard · OpenAI post · Vals Legal Research Bench
The configuration adds 5.8 points on LAB-AA, 1.6 on Harvey’s own benchmark and about 11 on Vals, a research test where a legal index should matter most. On Harvey’s LAB, by OpenAI’s numbers, most of the gain since GPT-6 Astra came from the base model.
LAB-AA: first place, at 20.8%
LAB-AA is the Artificial Analysis (AA) version of a benchmark from Harvey, the legal AI company whose Agent Builder we covered in April. It has 120 private assignments, developed by Harvey, across 24 practice areas. Under version 1.1, an assignment earns full credit only if the answer passes every rubric criterion and contains no material hallucination, meaning one that would mislead a reader on a substantive point. A panel of GPT-6 Sol, Grok 4.7 and Claude Opus 5.5 grades the criteria, and a separate GPT-6 Sol checker looks for hallucinations in two passes. AA reports no confidence intervals or run counts, and v1.1 scores are not directly comparable with v1.0. AA on LAB-AA v1.1 · OpenAI post
Selected rows from the leaderboard:
| Rank | Model (effort) | Full credit | Criteria passed | Material hallucinations per assignment | Cost per assignment |
|---|---|---|---|---|---|
| 1 | Astra for Law 6.2 (Max) | 20.8% | 94.4% | 0.03 | $7.33 |
| 2 | GPT-6.2 Astra (Max) | 15.0% | 93.8% | 0.06 | $6.35 |
| 3 | GPT-6.2 Astra (High) | 13.3% | 92.5% | 0.03 | $3.86 |
| 4 | Muse Wonder Full (Max) | 13.1% | 95.5% | 0.62 | not shown |
| 6 | GPT-6.2 Sol (Max) | 11.1% | 91.5% | 0.05 | $1.92 |
| 9 | Qwen4 Max | 9.4% | 91.2% | 0.36 | $7.85 |
| 11 | GPT-6 Astra (Max) | 8.6% | 90.6% | 0.03 | $13.67 |
| 13 | Claude Fable 5.2 (Max) | 7.2% | 91.0% | 0.75 | $39.84 |
| 21 | Claude Fable 5.2 (High) | 4.7% | 89.6% | 0.45 | $15.88 |
| 22 | Claude Opus 5.5 (Max) | 4.2% | 90.0% | 0.59 | $17.93 |
| 25 | Gemini 4 Argon (High) | 3.9% | 94.8% | 3.82 | $6.29 |
The best score on the board still means that about four in five assignments missed a requirement or contained a material hallucination. Crypto Briefing, a crypto-news outlet, made the same point in its coverage. We found no legal-press coverage of the November update. Crypto Briefing
Astra for Law does not pass the most individual criteria. Muse Spark 1.3 (96.0%), Muse Wonder Full (95.5%), Muse Wonder (95.1%) and Gemini 4 Argon (94.8%) all beat its 94.4%. It leads because of the hallucination gate: it had 0.03 material hallucinations per assignment against 0.62 for Muse Wonder Full and 3.82 for Gemini 4 Argon. In AA’s October data, which predates this model, more than 60% of otherwise-passing results across the models tested contained a material hallucination. For legal work, punishing fabrication harder than a thin answer seems to us the right weighting, but the ranking would look different without it.
OpenAI says Astra for Law earned full credit nearly 40% more often than the runner-up. That runner-up is its own base model, and in points the gap is 5.8. Against the best non-OpenAI entry, Muse Wonder Full, the lead is 7.7 points. With no intervals published, we can’t say how much of either gap would survive a rerun.
OpenAI’s headline hallucination claim is 95% fewer per assignment than Claude Opus 5.5: 0.03 against 0.59, a factor of about 20. The arithmetic holds, but Opus 5.5 is not Anthropic’s newest model, and on the leaderboard as we checked it several non-OpenAI rows are lower: Qwen4 Max at 0.36, and Claude Fable 5.2 at 0.45 (High), 0.53 (Low) and 0.56 (Medium). Against Fable 5.2’s lowest rate the factor is about 15, and against Qwen4 Max about 12. Astra for Law is still far below every non-OpenAI row. OpenAI cites AA’s latest comparison, which may be a different cut from the live table. AA tested Fable 5.2 at five effort settings, with a best full-credit score of 7.2%, and its Claude rows carry a label the page does not define. Fable 5.2 leads AA’s general Intelligence Index, with GPT-6.2 Astra second, so the legal ranking does not track the general one. LAB-AA leaderboard · AA on GPT-6.2
Other OpenAI rows share the 0.03 rate, including GPT-6 Astra at Max and GPT-6.2 Astra at High. Low hallucination looks like a trait of OpenAI’s recent models. At Max, the configuration took GPT-6.2 Astra’s rate from 0.06 to 0.03.
AA ran the evaluation, so this is not a vendor self-report, but all of the grading is by models. GPT-6 Sol, an OpenAI model, sits on the three-model panel and is also the hallucination checker. Crypto Briefing also notes that Harvey designed the benchmark, praised the upgrade and is set to receive API access to the model. Crypto Briefing
What the token numbers mean for cost
OpenAI says Astra for Law averaged about 64,000 output tokens per LAB-AA assignment against 139,000 for Muse Wonder Full, 54% less. We could not find those per-model counts in AA’s page text, and AA shows no cost for Muse Wonder Full, so the claim does not convert to dollars. OpenAI post
Fewer output tokens should mean a smaller bill, since GPT-6.2 Astra costs $40 per million output tokens against $8 for input. Those are the base model’s rates, and Astra for Law has no published price. The only dollar figures are the testers’ per-task costs, and they point in different directions. On LAB-AA the configuration costs about 15% more per assignment than GPT-6.2 Astra at Max ($7.33 against $6.35), and GPT-6.2 Astra at High reaches 13.3% for $3.86. On Vals it costs $2.75 per test on 77.74k tokens, where GPT-6.2 Astra costs $12.84 on 2.67 million. We did not confirm what Vals’s token figure includes. AA on GPT-6.2 · LAB-AA leaderboard · Vals Legal Research Bench
Vals and USCL are close races
On Vals AI’s Legal Research Bench, updated November 16, Astra for Law earned full credit on 112 of 208 test questions, 53.85%, which ranks sixth of 99 models. Gemini 4 Carbon leads at 55.77%. Muse Spark 1.3 Max, Claude Opus 5 and Claude Fable 5.1 share 55.29%, Gemini 4 Argon has 54.81%, and Gemini 4 Hydrogen ties Astra for Law. The gap to first is four questions. Vals shows no error bars on the leaderboard; by our arithmetic the standard error at this sample size is about 3.5 points, wider than the whole top group. OpenAI’s own wording, that Astra for Law “sits in the small cluster of top performers”, is fair, and the post does not claim first. OpenAI adds that the answers earned 90.9% of available points once partial credit counts. Vals’s judge is GPT 5.4, an OpenAI model. Vals Legal Research Bench
OpenAI also says the full-credit rate rose from about 46% in September. Vals’s page shows no September score, and the 54% that LawNext reported at launch was on a different 200-question validation set, so the eight-point gain rests on OpenAI’s statement alone.
Harvey published LAB: USCL on October 29. Models write briefs in 100 environments built from real federal dockets, and generative reward models compare the briefs in pairs against lawyer-reviewed criteria to produce Elo scores. OpenAI reports that when models had to find the cases themselves, Astra for Law scored 1,844 to Opus 5.5’s 1,817. With the record and cited opinions supplied, the two scored 1,883 and 1,882, which OpenAI calls “effectively even”. Harvey’s post predates GPT-6.2, names only Opus 5.5 as its top model and gives no per-model scores or intervals, so these figures come from OpenAI alone. Harvey does say nearly 700 Elo points separate its best and worst models, which makes 27 points a narrow lead. Harvey on LAB: USCL · OpenAI post
The post quotes Harvey’s head of legal research, Julio Pereyra, saying Harvey’s pairwise judges prefer the new briefs to GPT-6 Astra’s nearly 90% of the time. That compares the product with its predecessor, as judged by AI, not with attorneys.
Access, and what is not published
In September, LawNext reported that the Trusted Access Program targeted Am Law 200 firms, for lawyers and supervised staff. Qualifying firms got zero data retention on the API, and ChatGPT Enterprise use was excluded from human review by anyone at OpenAI. Boehmig said then that the zero-retention line is a one-sentence summary of an agreement of roughly 30 pages. OpenAI names Cooley, Ropes & Gray and Sullivan & Cromwell among the first firms. LawNext · Legal IT Insider · OpenAI post
Now any ChatGPT Enterprise customer can request access in the admin console, and teams that have it select gpt-6.2-astra-law in ChatGPT. It is the gated pattern OpenAI used for GPT-Rosalind in life sciences. Nothing we could open gives the approval criteria for a customer that is not a large law firm, says whether September’s retention and review terms carry over, or states a price. We could not confirm an API model ID either. OpenAI post
What to verify before relying on it
Jurisdiction coverage. Vals and USCL test U.S. law. The UK, Australian and Canadian additions and the 99.9% U.S. figure are OpenAI’s claims, and the 99.9% is limited to published precedential case law. Run questions from your own courts and practice areas, and test anything outside that scope separately.
Completeness as well as citations. As scored on LAB-AA, fabrication was rare, yet most assignments still fell short of full credit. Keep cite-checking, and also review output against the assignment’s requirements, since the low hallucination count implies that most lost credit came from unmet criteria.
The base model with your own retrieval. Legal-tech builders with a corpus and search of their own should test GPT-6.2 Astra against Astra for Law on the same tasks. The configuration’s lift ranged from 1.6 to about 11 points, and it was largest on research.
Terms in writing. Ask for the Trusted Access agreement and confirm retention, human review and API terms for your account.
Price per finished task. Until OpenAI publishes a rate, AA’s $7.33 and Vals’s $2.75 are rough indications. Get a quote before building a business case on either.
Sources
OpenAI
Benchmarks
- Harvey LAB-AA leaderboard, LAB-AA v1.1, GPT-6.2 Astra and GPT-6.2 Sol — Artificial Analysis
- Legal Research Bench — Vals AI
- Legal Agent Bench: U.S. Case Law — Harvey
Coverage
- OpenAI releases Astra for Law — LawNext
- OpenAI unveils Astra for Law — Legal IT Insider
- Astra for Law, September coverage — mixed-news
- OpenAI’s Astra for Law tops Harvey benchmark — Crypto Briefing
