Legal version of Astra appears in API model list
"gpt-6-astra-law," seen as the API version of "Astra for Law," the AI for law firms that OpenAI announced in September, appeared in the model list of its developer API. aitrackerbot, an X account that automatically detects and posts changes to AI services, reported it on Oct. 31 (5 p.m. the same day Japan time). OpenAI has not announced the start of API availability.
On the same day, Oct. 31, a "Legal model configurations" section was added to OpenAI's official API pricing page, listing gpt-6-astra-law. Pricing is $10 for input and $50 for output per million tokens, with cached input, a discount for repeatedly sent input, at $1. That is the same as the standard pricing for the general-purpose GPT-6 Astra ($10 for input, $50 for output and $1 for cached input). The pricing page notes that this is a promotional price through at least Dec. 31, 2026.
"gpt-6-rosalind," for life sciences, costs $15 for input and $75 for output, more than GPT-6 Astra. The legal version carries no premium. The pricing page does not say whether use of tools such as search is billed separately.
- Input
- Output
September announcement said API "soon"
OpenAI announced Astra for Law on Sept. 17 (Sept. 18 Japan time). It combines the top model "GPT-6 Astra" with tools, instructions and settings for legal research, rather than being a new model built from scratch. The company said it would be offered under the name "GPT-6 Astra Law" in ChatGPT and "gpt-6-astra-law" in the API.
Initially, availability was limited to vetted U.S. law firms, which could use it from ChatGPT and Codex. For the API, the company said it would be "available soon" without giving a timeline. It said legal AI companies such as Harvey and Legora would be able to build it into their own products through the API. For eligible firms, it will also offer an arrangement in which data sent to the API is not stored.
At its core is a dedicated search index of U.S. case law, statutes, regulations, court rules and administrative decisions. It contains more than 230 million URLs, with more added daily. Much of the case law is provided by the nonprofit Free Law Project, and it covers more than 99.9% of published U.S. precedential opinions, according to the company.
It is not clear whether general developers can use it right away. As of Nov. 3, OpenAI's help center still said API support was planned for the near future. The listing in the price list suggests that availability is near, but it is unknown whether the approval requirement has been lifted.
- OpenAI announces Astra for Law. API "available soon"Official
- Detection bot posts that gpt-6-astra-law appeared in API model listLeak
- Legal section added to API pricing page, with prices listedOfficial
- Help center still says API is planned for the near futureOfficial
OpenAI's measurement shows a 40% improvement in accuracy
According to OpenAI, on the 200-question validation set of the legal research evaluation "Legal Research Bench," Astra for Law correctly answered 54.0% of questions at its longest reasoning setting. GPT-6 Astra using only web search scored 38.7%, which OpenAI described as a 40% relative improvement. On case law questions, it found 24% more of the cases that should be cited, the company said. All of these are OpenAI's own measurements, and they have not been reproduced externally, such as by U.S.-based Vals AI, which created the evaluation. Even at 54.0%, the model failed to meet the standard for a correct answer on nearly half of the questions.
In its announcement, OpenAI showed an example in which Anthropic's "Claude Fable 5.1" returned a ruling that had been overturned on appeal. However, this was a single question-and-answer pair chosen by OpenAI, not a scored comparison.
General Astra ranks 34th in third-party evaluation
On the leaderboard for the same evaluation published by Vals AI (updated Oct. 29), the general-purpose GPT-6 Astra ranked 34th out of 89 models with an accuracy of 39.42%. Google's "Gemini 4 Carbon" led with 55.77%, followed by Meta's "Muse Spark 1.3 Max" and Anthropic's Claude Opus 5 and Claude Fable 5.1 at 55.29%. OpenAI's higher-tier model "GPT-6.1 Sol" scored 38.46%. Astra for Law itself is not on the leaderboard.
The leaderboard is measured on a private test set of 208 questions, separate from the validation set OpenAI used. Because the questions differ, Astra for Law's 54.0% cannot be compared directly with the leaderboard figures. Once it becomes generally available, a key question will be whether it can be compared with other companies' models under the same conditions on Vals AI's leaderboard.
The leaderboard gives each model the same five tools, such as case law search and web search, and uses AI grading to calculate the share of questions that meet all grading criteria.
Cost per question higher than top-ranked Gemini
Vals AI also publishes the API cost of solving each question. GPT-6 Astra cost $10.58 per question, more than the $8.74 for top-ranked Gemini 4 Carbon. Claude Opus 5 cost $6.76 and scored 55.29%, and Meta's Muse Spark 1.3 Max cost $0.59 and also scored 55.29%. On the other hand, Anthropic's Claude Fable 5.1 cost $23.06 and Claude Opus 5.5 cost $25.16, more than twice GPT-6 Astra. The legal version is not on the leaderboard, and its cost per question has not been published.
Major players pile into legal AI
Major companies keep entering the market for AI for law firms. According to Reuters, Anthropic has offered legal tools since January, and Google expanded legal features in its business-oriented Gemini Enterprise in August. U.S. legal data giant Thomson Reuters also released "Thomson 1.0," a model trained on its own legal materials, in August. The company is also a partner for Astra for Law.
Markets have reacted sharply to moves in legal AI. When Anthropic released a legal plugin in February, shares of Thomson Reuters fell 14.2%, Britain's RELX fell 14% and the Netherlands' Wolters Kluwer fell 10.5%. Harvey, which is said to be able to build Astra for Law into its products, was valued at $15.6 billion in September, and Legora is reported to be seeking funding at a $10 billion valuation.
