OpenAI Lists Astra for Law at $10/$50 - No Premium for Lawyers

For six weeks, OpenAI's legal AI configuration has been a product with partners, a benchmark, and no price. OpenAI introduced Astra for Law on September 17, 2026, named six major law firms and two AI companies building on it, and said the API would arrive as gpt-6-astra-law, but published no rate. On October 31, the price finally appeared, quietly: a new "Legal model configurations" table on OpenAI's own API pricing page lists gpt-6-astra-law at $10.00 per million input tokens and $50.00 per million output tokens, the exact rate the page gives GPT-6 Astra itself. This article walks through what the page shows, how the three specialist configurations compare, and the fine print behind each number.

What the pricing page actually shows

The listing was first spotted on October 31 by TestingCatalog, which noted on X that gpt-6-astra-law was "now listed on the API pricing page." FourWeekMBA read the page the same morning and reported the figures; the page itself was then read directly for this article on October 31, 2026, and every number below comes from that live read.

The page's new section, headed "Legal model configurations" and described as "our latest legal model configurations," contains one row:

Directly under the table, the page states: "GPT-6 Astra for Law promotional pricing is available at least through December 31, 2026." Read that sentence carefully. It sets a floor date for the current price (the rate holds at least through year-end), not a ceiling after which a new price is announced. The page says nothing about what the configuration costs after the promotional period. That is an honest reading of the page, and worth holding onto before projecting budgets past January.

The law table sits alongside two other specialist tables that make the comparison interesting. The "Life sciences models" table, described as "our latest Rosalind models," lists gpt-6-rosalind at $15.00 input, $75.00 output, and $1.50 cached input per million tokens. The same table prices gpt-rosalind-research and gpt-rosalind-discovery at $5.00 input and $25.00 output each. These are cheaper Rosalind variants, but the page limits all Rosalind access to "approved internal research through the trusted-access program." The "Cyber models" table, described as "our latest Daybreak models," lists gpt-5.6-cyber at $12.50 input, $75.00 output, and $1.25 cached input, against gpt-5.6-sol at $4.00 input and $20.00 output in the same table.

The three specialist configurations, compared

The arithmetic is straightforward, done from the page's own figures:

The striking row is the first one. A specialist vertical configuration, carrying a legal search index and legal-analysis instructions, costs exactly what the base flagship costs: 1.0x on input, 1.0x on output, 1.0x on cached input. The life-sciences line costs 1.5x GPT-6 Astra on both input and output. The cyber line costs 3.1x the input price of gpt-5.6-sol and 3.75x its output price. Lawyers, in other words, face no premium for the specialization; life-sciences researchers and security teams do.

For UK and Canadian readers, all three tables bill in US dollars. At October 31, 2026 market rates (about £0.747 and C$1.397 per dollar, verified at publication), the law configuration costs roughly £7.47 or C$13.97 per million input tokens and £37.35 or C$69.85 per million output tokens. Rosalind works out to roughly £11.21/C$20.96 input and £56.03/C$104.78 output; the cyber model to roughly £9.34/C$17.46 input and £56.03/C$104.78 output. These are close approximations, not quotes — your card provider converts at its own rate.

What Astra for Law actually is

The first misunderstanding to clear up: this is not a new model. As OpenAI described it on September 17, Astra for Law "combines GPT-6 Astra, our latest and most powerful model, with settings, tools, and context tailored for professional legal work." It appears in the ChatGPT model picker as "GPT-6 Astra Law" and, when the API opens, as gpt-6-astra-law. What is new underneath is the index, the instructions, and the integrations.

The legal search index covers more than 230 million URLs of U.S. case law, statutes, regulations, court rules, and administrative decisions, with sources added daily, according to OpenAI. The case-law portion has a named source: the company's work with the Free Law Project, the nonprofit behind CourtListener, brings a collection covering more than 99.9 percent of published U.S. precedential case law into the research experience. That percentage is CourtListener's claim about its own database; Free Law Project's coverage page states the same figure. It is not a measurement OpenAI performed. It is also a bounded one: published precedential opinions only, so unpublished and non-precedential decisions sit outside it.

The launch benchmark is worth reading with the same care. OpenAI tested Astra for Law on 200 questions from the private validation set of Vals AI's Legal Research Bench and reported that, at the highest reasoning effort for both systems, the configuration passed the overall correctness check on 54.0 percent of questions, against 38.7 percent for GPT-6 Astra using web search alone, a 40 percent relative improvement. It also reported finding 24 percent more reference cases on case-law questions and retrieving up to 54 percent more relevant passages from the correct opinions. As independent analysis of the launch noted, those gains are measured against the same model with plain web search: the index clearly adds something, but OpenAI ran the test itself, the questions are not public, and no competing legal product appears in the measurement. The announcement also includes two hand-picked worked examples against Claude Fable 5.1: OpenAI narrating its own prompts, not benchmark data.

The launch came with 26 partner-built plugins connecting ChatGPT to tools firms already run (Relativity, Clio, iManage, Intapp, DeepJudge, and Thomson Reuters among them), plus nine community plugins from lawyers and legal engineers. Named users include Sullivan & Cromwell, Ropes & Gray, Cooley, Skadden, Latham & Watkins, and Wachtell, Lipton, Rosen & Katz, with several having built their own tools on the stack: an agreement analyzer, a diligence system, a capital-markets drafting tool. Harvey and Legora, which sell their own legal AI systems, are named as API customers cleared to build on the configuration inside their own products.

The fine print that matters

Start with what the pricing page does not establish. It lists the configuration and its price. It does not include a separate announcement saying gpt-6-astra-law is generally available on the API, or that the Trusted Access limits for law firms have changed. As of the September 17 announcement, the product was "initially offered to selected law firms through Trusted Access in ChatGPT and Codex," with the API "coming soon." A price on the page is a strong signal that the API is getting closer, but it is not itself an availability announcement.

The promo line cuts the other way too. "Promotional pricing is available at least through December 31, 2026" guarantees the rate through year-end; it guarantees nothing about January. A firm modeling a multi-year legal AI budget should treat $10/$50 as the near-term rate and ask OpenAI about the post-promo structure in writing.

Two page-wide notes apply to the law price as well. Regional processing endpoints carry a 10 percent uplift for models released on or after March 5, 2026, and starting November 1, 2026, FedRAMP endpoints carry the same uplift. Both add to the token rates above for eligible workloads.

On access terms, OpenAI says eligible firms get Zero Data Retention on the API (requests are not kept) and that ChatGPT Enterprise usage is excluded from human review by default. Those terms are aimed at the confidentiality objection courts have raised: Judge Stephen Alexander Vaden of the U.S. Court of International Trade has noted that generative systems "create novel risks to the security of confidential information," since prompt material can be retained by the model owner. Separately, Judge Nina Y. Wang of the U.S. District Court for the District of Colorado requires an AI certification on filings, with each filer certifying that citations reference real cases. The lesson: the lawyer, not the model, answers for the filing.

One more boundary: the index is a U.S. legal corpus. Firms practicing in other jurisdictions get the model and the plugins but not a native corpus for their own law. That is a genuine limit for the UK and Canadian readers of this site.

How this fits the legal-AI race

Astra for Law is not the first lab into this vertical; it is the third in five months. Anthropic launched Claude for Legal on May 12, 2026, and Google launched Gemini Enterprise for Legal on August 25. The same incumbent legal-tech vendor, Thomson Reuters, has wired its CoCounsel Legal integration into both Claude and ChatGPT on nearly identical terms — a sign that no single lab has locked up the market.

There is a second tension worth naming. Harvey, the standalone legal-AI vendor OpenAI's own announcement lists as an API customer cleared to build on Astra for Law, raised at a $15.5 billion valuation eight days before the September launch. OpenAI is simultaneously treating Harvey as a partner and shipping a first-party product, sold to the same law firms, into the market Harvey raised over a billion cumulative dollars to serve. For Harvey and Legora, the $10/$50 rate is the wholesale price of the engine inside their own products.

The pricing strategy itself is the real story. Charging 1.0x for the legal configuration, with no premium for the index, the instructions, or the plugins, reads as a land grab: price the specialist surface at the commodity rate and take share from incumbents while the competition is measured in months, not years. Rosalind at 1.5x and the cyber model at 3.1x show OpenAI knows how to charge for specialization. For law, it chose not to — at least through December 31.

Pro tips and common mistakes

Treat the price as near-term, not permanent. "Promotional pricing is available at least through December 31, 2026" is a floor date. Budget the first quarter on $10/$50 and get the post-promo structure from OpenAI before committing a multi-year matter pipeline.

Do not confuse the listing with availability. The pricing page lists the configuration; OpenAI has not announced general API access. If your firm is not in the Trusted Access program, the row on the pricing page does not yet buy you anything.

Price the cached-input row, not just input/output. At $1.00 per million cached tokens against $10.00 fresh, a retrieval-heavy legal workflow that re-reads the same authorities pays meaningfully less than the headline rate suggests. The 90 percent discount on cache reads is the quiet line in the table.

Watch the benchmark's limits. 54.0 percent overall correctness is OpenAI's own measurement against its own prior setup on non-public questions. The Vals public leaderboard, on a different question set, places rival models at 55.29 percent strict all-pass accuracy. Ask for cross-lab numbers on the public benchmark family before treating any vendor's private figure as a purchasing decision.

Remember who signs the filing. Court certification rules place the citation-accuracy burden on the filing lawyer. A 40 percent relative improvement changes how much checking is needed, not whether checking is needed.

Frequently asked questions

How much does gpt-6-astra-law cost per million tokens?

OpenAI's API pricing page lists $10.00 per million input tokens, $50.00 per million output tokens, and $1.00 per million cached input tokens, identical to GPT-6 Astra itself. The rate is promotional, available at least through December 31, 2026.

Is Astra for Law a separate AI model from GPT-6 Astra?

No. OpenAI describes it as a configuration: GPT-6 Astra combined with a legal search index of more than 230 million URLs, legal-analysis instructions, and plugins. The model weights are the same; the search corpus and tooling around it are what is new.

Can anyone call gpt-6-astra-law on the API today?

Not yet. The pricing page lists the configuration, but OpenAI has not announced general API availability. Access remains through the Trusted Access program for selected law firms in ChatGPT and Codex, with the API described as coming.

How do the other specialist models compare on price?

The life-sciences model gpt-6-rosalind costs $15/$75 per million tokens, 1.5x GPT-6 Astra, and is limited to approved internal research through the trusted-access program. The cyber model gpt-5.6-cyber costs $12.50/$75, or 3.1x the input price of gpt-5.6-sol.

Does Astra for Law cover non-U.S. law?

No. The legal search index covers U.S. case law, statutes, regulations, court rules, and administrative decisions. Firms in the UK, Canada, or elsewhere get the model and plugins but no native corpus for their own jurisdiction.

OpenAI has priced its legal AI like a land grab: the same rate as the flagship, with a year-end promo floor. For context on how OpenAI prices the rest of its frontier lineup, see our GPT-6.1 Terra pricing breakdown and our coverage of the GPT-6 launch on Google Cloud, plus how the GPT-6.1 Sol reasoning model is priced. On the buyer side, Anthropic's discounted Claude deal for government shows how the other lab prices regulated-sector access.

Sources

• OpenAI, API pricing page — read live October 31, 2026 (legal, life-sciences, and cyber tables; promotional pricing line; regional and FedRAMP uplifts)
• FourWeekMBA, OpenAI Lists Astra for Law at GPT-6 Astra's Own Price — October 31, 2026 (discovery of the listing via TestingCatalog)
• YFarmX, Astra for Law: what OpenAI's legal configuration actually measures — September 2026 (announcement details, the two 54% figures, Vals leaderboard context, court orders)
• Recul, Astra for Law: OpenAI puts US case law inside ChatGPT — September 2026 (configuration details, 26 plugins, named firms, access terms)
• Ena Pragma, Astra for Law is real. The benchmark a buyer actually needs is not on the page. — September 20, 2026 (three-labs legal race, Harvey valuation tension, sanctions-risk scale)

Related articles

• GPT-6.1 Terra: OpenAI's $1 Computer-Use Model, 3x Faster Than Astra
• OpenAI Comes to Google Cloud: GPT-6 Models and Codex Land in Model Garden
• OpenAI Unveils Next-Generation Reasoning Models
• Anthropic Ends $1 Claude for Government Today; 40% Off Next

This is general information, not procurement or legal advice.