Quick answer
On September 17, 2026, OpenAI released Astra for Law: GPT-6 Astra combined with a dedicated legal search index and profession-specific instructions. In OpenAI's own evaluation it answered legal research questions with 54.0% overall correctness, against 38.7% for GPT-6 Astra using ordinary web search. Pricing has not been published. It is the first time a frontier lab has shipped a version of its flagship model configured for a single profession.
Legal AI has been one of the most valuable applications of language models, and until now it belonged to specialists: Harvey, Thomson Reuters' CoCounsel, Lexis+ AI, and a crowd of startups building on top of OpenAI and Anthropic models. Astra for Law is OpenAI stepping into that layer itself. The interesting question is not whether the model is good — it is what happens to the companies whose business was making OpenAI's models good at law.
What it actually is
- GPT-6 Astra, the same model released September 3, with no separate training run announced
- A dedicated legal search index — case law, statutes, regulations — the model retrieves from instead of the open web
- Custom instructions that shape how it reasons about and cites legal sources
- A 15-point jump in correctness on OpenAI's legal evaluation, from 38.7% to 54.0%
- Pricing and availability details not yet published
Read the 54% carefully
Fifty-four percent correctness is a big improvement and still a coin flip. That is roughly the level where a tool is useful for a lawyer who checks every answer and dangerous for anyone who does not. The number also comes from OpenAI's own benchmark, which it has not described in detail. The right comparison is not Astra for Law against web search; it is Astra for Law against Harvey or CoCounsel on the same questions, and no one outside those companies has run it.
What it means for legal AI companies
- The pure "GPT plus legal prompt" products are in trouble — OpenAI now does that itself
- Products with proprietary content, such as Thomson Reuters and LexisNexis, keep their moat: the index is the value, and theirs are deeper
- Workflow products — Harvey's document review pipelines, Spellbook inside Word, Ironclad's contract lifecycle — compete on integration, not on the model
- Expect Astra for Law to appear inside those products as a model option rather than replacing them outright
This is the wrapper debate made concrete. If your legal product's value was a prompt and a model, OpenAI just shipped it. If it was data, workflow, and trust, you still have a business — and a better model to build it on.
Related reading
Bottom line
Astra for Law is a signal more than a product for now: pricing is unknown, the benchmark is internal, and 54% is not a number to trust unsupervised. The signal is that frontier labs will ship vertical configurations of their models directly. Legal is first because it is lucrative and text-heavy. Medicine, finance, and tax will follow.
