OpenAI released an updated Astra for Law on Tuesday, 17 November, and opened it to every ChatGPT Enterprise customer through its Trusted Access Program, including API access, according to its announcement. OpenAI describes Astra for Law as a model configuration of GPT-6.2 Astra that pairs the model with “an expanded legal search index and instructions for researching a question, weighing the authorities, and writing the answer”.
Our number: on Vals AI’s Legal Research Bench, Astra for Law costs $2.75 per test against $12.84 for plain GPT-6.2 Astra, 79% less, while passing every check on 53.85% of questions against 42.79%, 11.1 points higher, by our arithmetic from the Vals leaderboard. The legal setup makes the same base model both cheaper and more accurate on that test.
Business Pill · Usage-Based Pricing
An evergreen Business Engineer explainer on usage-based pricing, paying for work rather than seats. It teaches the general idea only and does not discuss OpenAI or Astra for Law.
The key insight: OpenAI’s legal product is the same GPT-6.2 Astra with a search index and written instructions on top, and on Vals AI’s Legal Research Bench that setup costs $2.75 per test against $12.84 for the bare model while passing 11.1 points more questions, by our arithmetic.
What OpenAI Released
OpenAI says the update is “state-of-the-art on Harvey LAB-AA and a top performer on LAB: USCL”, a benchmark co-developed by Harvey and Grubel, and on Vals AI’s Legal Research Bench. OpenAI first “introduced Astra for Law in September” to a small group of firms, including Cooley, Ropes & Gray and Sullivan & Cromwell.
OpenAI says it grew the legal search index by 34% since September, “adding over 80 million new URLs”, including the UK, Australia and Canada. It says the index now spans over 310 million URLs and covers “more than 99.9% of published U.S. precedential case law”.
Enterprise customers request it in the admin console; teams with access select “gpt-6.2-astra-law” in ChatGPT, and API access goes through a contact form, OpenAI says. The X account @TheRealAdamG posted the release at 18:19 UTC. OpenAI did not publish an API price for the configuration in the post.

What the Vals Leaderboard Shows
Vals AI’s leaderboard, updated 16 November, ranks Astra for Law sixth at 53.85%. Gemini 4 Carbon is first at 55.77%, followed by Muse Spark 1.3 Max, Claude Opus 5 and Claude Fable 5.1 at 55.29% each, and Gemini 4 Argon at 54.81%. OpenAI calls the group “the small cluster of top performers”.
Vals lists Astra for Law’s cost per test at $2.75, against $8.74 for Gemini 4 Carbon, $6.76 for Claude Opus 5 and $23.06 for Claude Fable 5.1. Muse Spark 1.3 Max is cheaper at $0.59. Among the top six, Astra for Law is second-cheapest, and Gemini 4 Carbon costs 3.2 times as much per test, by our arithmetic.
OpenAI says Astra for Law’s full-credit share on Vals rose from about 46% in September to 53.8% in Vals’ 16 November evaluation, “112 of 208 questions”. Plain GPT-6.2 Astra sits 31st on the same leaderboard.
What OpenAI Claims on Harvey’s Tests
On LAB-AA, Artificial Analysis’s version of Harvey’s Legal Agent Benchmark with 120 assignments across 24 areas of law, OpenAI says “Astra for Law ranked first, earning full credit on 20.8% of assignments, ahead of GPT‑6.2 Astra at 15.0% and Muse Wonder Full at 13.1%”. That is 39% more often than GPT-6.2 Astra, the “nearly 40%” in its summary, by our arithmetic.
OpenAI says Astra for Law averaged about 64,000 output tokens per assignment, compared with 139,000 for Muse Wonder Full. Divided by each full-credit rate, that is roughly 308,000 output tokens per fully passed assignment against about 1.06 million, about 3.4 times fewer, by our arithmetic from OpenAI’s figures.
On hallucinations, OpenAI says Astra for Law averaged 0.03 material hallucinations per assignment, “compared with 0.59 for Claude Opus 5.5”. On Harvey’s own LAB, it says the full-credit share went from 4.6% for GPT-6 Astra to 14.6% for GPT-6.2 Astra and 16.2% with the legal configuration.
Harvey’s head of legal research, Julio Pereyra, is quoted in the post saying the new briefs “are preferred to those of GPT‑6 Astra nearly 90% of the time by our pairwise judges”. On LAB: USCL, OpenAI says Astra for Law scored 1,844 to Opus 5.5’s 1,817 when the models did their own research, and 1,883 to 1,882 when given the precedents.
The Business Lever
The lever is ChatGPT Enterprise: OpenAI moved the configuration from a small group of firms to every Enterprise customer, with legal API access by request. The pitch it makes is cost: OpenAI says lower output “can mean the difference between staying within budget and blowing through it” for teams paying by usage.
It follows OpenAI’s GPT-6.2 Astra release earlier on Tuesday, and lands in a legal market where AI-native firms such as Alecto, which raised a $20M seed, are also selling legal work.
The Structural Read
As we read the Vals numbers, the configuration matters more than the model on this test: the same GPT-6.2 Astra moves from 31st to sixth on the leaderboard with the legal index and instructions added, at about a fifth of the cost per test.
As we read the access change, OpenAI is selling the configuration through ChatGPT Enterprise rather than as a separate product: customers request it in the admin console and select a model id, while API access goes through a sales form with no published price.
OpenAI, 17 November 2026
“Building Astra for Law as a model configuration lets us stay on the bleeding edge of the frontier”
Three Implications
FOR LAW FIRMS AND LEGAL TEAMS Every ChatGPT Enterprise customer can now request Astra for Law through the Trusted Access Program, OpenAI says; since September it had been with a small group of firms, including Cooley, Ropes & Gray and Sullivan & Cromwell.
FOR MODEL BUYERS On Vals AI’s leaderboard Astra for Law ranks sixth at 53.85% and $2.75 per test; the five models above it score between 54.81% and 55.77% at $0.59 to $23.06 per test.
FOR HARVEY Harvey co-developed LAB and LAB: USCL and its head of legal research is quoted in OpenAI’s post; the 20.8% LAB-AA result OpenAI reports was not in the summary of Artificial Analysis’s public page we read.
The Business Engineer Lens
This story maps onto the Business Engineer framework AI & The Harness Theory.
The framework says of the harness: “It is an operating system that wraps your AI — a harness that compresses your thinking, your team’s feedback, and the model’s raw capability into something replicable.”
As we read it, Astra for Law is that idea sold as a product: OpenAI says it refines the configuration’s settings and instructions in response to practitioner feedback, on top of GPT-6.2 Astra. The framework predates this release and does not discuss OpenAI or Astra for Law.
What Is Not Established
The LAB-AA, LAB and LAB: USCL results are as reported by OpenAI. The default view of Artificial Analysis’s LAB-AA page we read names GPT-6.2 Astra (Max) highest at 15.0%; it shows 27 of 44 models and has a filter for models with no public API, and Astra for Law was not in that summary. The Vals figures come from Vals’ own leaderboard.
OpenAI did not give a price for Astra for Law, so Vals’ cost per test is Vals’ measurement, not a list price. We did not contact OpenAI, Harvey or Vals AI.
The Bottom Line
OpenAI opened Astra for Law to all ChatGPT Enterprise customers on 17 November. On Vals AI’s Legal Research Bench it ranks sixth at 53.85%, 11.1 points above plain GPT-6.2 Astra, at $2.75 per test against $12.84, by our arithmetic. Its first-place LAB-AA result of 20.8% is OpenAI’s report.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Release details, access, index figures and the LAB-AA, LAB and LAB: USCL results are from OpenAI’s post of 17 November 2026; the Harvey quote is as published there. Vals AI ranks, pass rates and costs per test are from Vals AI’s Legal Research Bench leaderboard, updated 16 November 2026. The LAB-AA page summary is from Artificial Analysis. The cost and ratio comparisons are our arithmetic. We did not contact OpenAI, Harvey or Vals AI.
Sources: OpenAI, Astra for Law leads in legal intelligence, 17 Nov 2026 · Vals AI, Legal Research Bench leaderboard, updated 16 Nov 2026 · Artificial Analysis, Harvey LAB-AA v1.1 · @TheRealAdamG on X, 17 Nov 2026









