OpenAI released Astra for Law on September 17

OpenAI released Astra for Law on September 17. It is not a new model: it is GPT-6 Astra configured for legal work, wrapped in a legal search index and a set of instructions written for legal analysis.

The difference sits exactly there: standard GPT-6 Astra looks at general web search, Astra for Law looks at its own legal index.

The index reaches more than 230 million URLs: US case law, statutes, regulations, court rules and administrative decisions, with sources added daily.

The case law comes from CourtListener, run by the nonprofit Free Law Project. OpenAI says that collection covers more than 99.9% of published precedential US case law. The most critical layer under the product is not OpenAI’s own data, it is a civil society archive.

The measurement OpenAI published rests on 200 questions from Vals AI’s Legal Research Bench. At the highest reasoning setting, Astra for Law passed the correctness check on 54% of them, while GPT-6 Astra with web search alone stopped at 38.7%.

The answers also got longer: roughly twice the length of the web-only version on average, with 24% more case law citations.

The chart OpenAI published shows the number has a price. The headline 54% comes from the most expensive setting, roughly $5.67 per answer.

Two settings down gives 52.8% for $3.49 (the setting in between gives 53.2% for $3.96). So the last 1.2 points cost more than 60% extra. GPT-6 Astra with web search peaks at $3.38 and then slips back at a higher price.

OpenAI published a head to head against a direct competitor. Given the same legal memo prompt, Astra for Law found two closely matching precedents.

According to OpenAI, Claude Fable 5.1 cited a judgment that had been overturned on appeal in the litigation example, and reported that it could not find such a case in the transactional one. OpenAI picked the examples and ran the comparison.

Access is limited to selected firms. Those in the Trusted Access program reach it inside ChatGPT and Codex, where it shows up in the model picker as GPT-6 Astra Law. Latham & Watkins, Ropes & Gray, Sullivan & Cromwell and Skadden are among the firms worked with.

Eligible firms get zero data retention on the API. Harvey and Legora are API customers. The gpt-6-astra-law API comes later, with no date and no pricing.

54% also means this: it fails the correctness check on 46% of the questions. A fifteen point gain changes how much checking is needed, not whether checking is needed.

The number is OpenAI’s own measurement, taken on a private 200 question set, and no independent lab has reproduced it. Astra for Law does not appear at all on Vals AI’s public leaderboard; in the September 3 snapshot the top score there is Claude Opus 5 at 55.29%. The two are not directly comparable, but the claim has no public counterpart.

Hallucination in law is an old problem. A 2024 Stanford study found that even purpose-built legal tools hallucinated in 17% to 33% of their answers. That study did not measure Astra for Law, but it sets the floor.

OpenAI’s own announcement implies the same: what it promises is not a finished answer but a research foundation built on authorities the lawyer can review personally.