OpenAI Walked Into Harvey's House and Brought a Gift
OpenAI introduced Astra for Law on September 17: GPT-6 Astra wired to a legal search index of more than 230 million URLs, plus the access controls and workflow tooling that law firms actually require before anything touches a client matter. The index covers US case law, statutes, regulations, court rules, and administrative decisions, with sources added daily. Announcement at https://openai.com/index/astra-for-law/
Rollout is deliberately narrow at first. Selected firms get it through Trusted Access in ChatGPT and Codex, with API availability to follow. It launches with 26 partner plugins from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge, and iManage, and OpenAI explicitly names Harvey and Legora as API customers who will build on Astra for Law.
Read that last sentence again, because it is the whole story. Harvey and Legora are the two best-funded legal AI companies on earth, and their model supplier just shipped a legal-configured model with a legal search index and firm-grade access controls under its own brand. They are named as partners. They are also now building on a floor that their competitor sets, in a category their competitor has decided to enter by name. Vertical AI startups have spent three years betting that domain workflow and trust are the moat and the model is a commodity. This is the test of that bet, in the single most lucrative vertical.
The piece that is not a commodity here, and the reason this is more than a system prompt with a suit on, is the index. Two hundred thirty million URLs of primary legal source material, refreshed daily, is an asset with real acquisition and maintenance cost, and it is the thing a startup cannot replicate over a weekend. The retrieval corpus is the product. The model is the delivery vehicle.
What nobody has published yet is a benchmark. There is no accuracy number against a professional standard, no hallucinated-citation rate, no comparison to the incumbents on a shared task. In a field where fabricating a case citation gets lawyers sanctioned by actual judges, that absence is the number everyone should be asking for. Same lab, same week, [OpenAI published an unflattering misalignment finding](https://clauday.com/article/7a5cfafb-5102-4652-9729-ac68465f78fe) about a model in the same family.
← Back to all articles
Rollout is deliberately narrow at first. Selected firms get it through Trusted Access in ChatGPT and Codex, with API availability to follow. It launches with 26 partner plugins from Thomson Reuters, Intapp, Harvey, Legora, DeepJudge, and iManage, and OpenAI explicitly names Harvey and Legora as API customers who will build on Astra for Law.
Read that last sentence again, because it is the whole story. Harvey and Legora are the two best-funded legal AI companies on earth, and their model supplier just shipped a legal-configured model with a legal search index and firm-grade access controls under its own brand. They are named as partners. They are also now building on a floor that their competitor sets, in a category their competitor has decided to enter by name. Vertical AI startups have spent three years betting that domain workflow and trust are the moat and the model is a commodity. This is the test of that bet, in the single most lucrative vertical.
The piece that is not a commodity here, and the reason this is more than a system prompt with a suit on, is the index. Two hundred thirty million URLs of primary legal source material, refreshed daily, is an asset with real acquisition and maintenance cost, and it is the thing a startup cannot replicate over a weekend. The retrieval corpus is the product. The model is the delivery vehicle.
What nobody has published yet is a benchmark. There is no accuracy number against a professional standard, no hallucinated-citation rate, no comparison to the incumbents on a shared task. In a field where fabricating a case citation gets lawyers sanctioned by actual judges, that absence is the number everyone should be asking for. Same lab, same week, [OpenAI published an unflattering misalignment finding](https://clauday.com/article/7a5cfafb-5102-4652-9729-ac68465f78fe) about a model in the same family.
Comments