OpenAI is opening Astra for Law, its legal configuration of GPT-6.2 Astra, to all ChatGPT Enterprise customers, the company said on 17 November; organisations can request access through the Trusted Access Program in the admin console; API access is also available. On the LAB-AA benchmark, the new version earned full credit on 20.8% of the 120 legal assignments, more than any other model tested, according to OpenAI.
Key facts
- Since 17 November 2026, every ChatGPT Enterprise customer can request Astra for Law through OpenAI's Trusted Access Program, including API access.
- On LAB-AA's 120 assignments, Astra for Law earned full credit on 20.8%, against 15.0% for GPT-6.2 Astra, according to OpenAI.
- All cited tests use US law; OpenAI says the index covers more than 310 million URLs and does not mention German law.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
Astra for Law is not a separate model. It pairs GPT-6.2 Astra with a legal search index and fixed instructions for researching a question, weighing the authorities and writing the answer. OpenAI first offered it in September to a small group of firms including Cooley, Ropes & Gray and Sullivan & Cromwell. In ChatGPT the model now appears as gpt-6.2-astra-law. OpenAI does not give a price in the announcement.
The search index has grown by 34% since September, OpenAI said, adding more than 80 million URLs, including from the UK, Australia and Canada. It now spans more than 310 million URLs and over 99.9% of published US precedential case law. The post does not mention German, Austrian or Swiss law.
Astra for Law benchmark results and limits
The figures come from OpenAI and refer to two tests run by outside providers:
| Test | Astra for Law | Comparison |
|---|---|---|
| LAB-AA (Artificial Analysis, 120 assignments across 24 areas of law): full credit | 20.8% | GPT-6.2 Astra 15.0%, Muse Wonder Full 13.1% |
| LAB-AA: material hallucinations per assignment | 0.03 | Claude Opus 5.5: 0.59 |
| LAB-AA: output tokens per assignment | about 64,000 | Muse Wonder Full: about 139,000 |
| Vals AI Legal Research Bench (208 US-law questions): full credit | 53.8% (112 questions) | September: about 46% |
On LAB-AA, an answer only earns full credit if it meets every requirement of the assignment and makes no material claim the source documents do not support. The assignments were written by Harvey, a legal AI software company, and range from deposition outlines to commercial lease reviews. The Legal Research Bench questions are set by practising lawyers on US federal and state law. Counting partial credit, Astra for Law earned 90.9% of the available points there, OpenAI said.
OpenAI presents the shorter output as a cost advantage for legal teams that pay by usage. The figures have not yet been independently reproduced; the table reflects the vendor's account. A different picture of how models compare on real work came from Epoch AI's test of Fable 5.2 and GPT-6.2 Astra. OpenAI itself flagged weaknesses of the base model at launch, covered in GPT-6.2 Astra is 20% cheaper but hides problems more often.
The move fits a run of industry products OpenAI has launched this week, most recently ChatGPT for Sales with more than 30 connected sales tools. AI agents are also taking on audit work, as KPMG's Clara intelligence in the US shows.
Analysis
OpenAI's headline says the product leads. The number behind it is more sober: the best model in the test completes roughly one legal assignment in five in full and without a material error. On the research questions, almost half of the answers miss at least one required point. Astra for Law has improved clearly since September, but its output does not belong in a brief or contract without review.
For legal departments and law firms in Germany, Austria and Switzerland there is a further gap: every test cited uses US law, and OpenAI says it is expanding the index into English-language jurisdictions. Nothing in the post shows whether the low hallucination rates hold for German tenancy law or Swiss employment law. A sensible step is a sample of closed matters from a firm's own practice before an admin approves access, plus a fixed rule on who checks every citation against the original source. Responsibility stays with the person who signs the work, so approving access in the admin console is also a question of organisation.
Got AI news? Submit news →
FAQ
What is Astra for Law?
A legal configuration of GPT-6.2 Astra that combines the model with a legal search index and fixed research instructions. In ChatGPT it appears as gpt-6.2-astra-law.
Who can use Astra for Law?
Since 17 November 2026, all ChatGPT Enterprise customers can request access through the Trusted Access Program in the admin console, and API access via a form. OpenAI gives no price.
Does Astra for Law cover German law?
OpenAI gives no information on this. The published tests use US law, and the index has been expanded to the UK, Australia and Canada.
Source
OpenAI: Astra for Law leads in legal intelligence
More sources
- Vals AI – Benchmark-Anbieter, Werte laut OpenAI-Beitrag (Auswertung vom 16.11.2026).
- The Neuron – Hinweisgeber, nicht Quelle.




