On this page · 11 sections
- What OpenAI actually lists
- What the legal configuration is
- Who can use it, and how
- What the benchmarks say, from three sources
- A worked cost model
- What it means for law firms
- What it means for legal-tech vendors
- Data residency and India-specific considerations
- A six-step plan before 31 December
- How eCorpIT can help
- References
Summary. OpenAI's API pricing page, checked on 20 November 2026, lists gpt-6.2-astra-law at $8.00 per million input tokens and $40.00 per million output tokens, with cached input at $0.40, under a heading for legal model configurations. Requests above 272,000 input tokens cost $16.00 and $60.00 instead. The older gpt-6-astra-law is listed at $10.00 and $50.00. The page says this promotional pricing runs at least through 31 December 2026, 41 days from today, and publishes no rate for afterwards. Today the legal configuration costs exactly the same per token as plain GPT-6.2 Astra, which OpenAI added to the API on 17 November 2026. The same day, OpenAI said Astra for Law now runs on GPT-6.2 Astra, opened access requests to every ChatGPT Enterprise customer through its Trusted Access Program, and invited API applications. On Artificial Analysis's Harvey LAB-AA test it scored 20.8% against 15.0% for GPT-6.2 Astra, while Vals AI's Legal Research Bench ranks it seventh at 53.85%. Our verdict: apply now and test on your own matters while there is no premium, but budget 2027 as if the price could rise, and keep a fallback to the base model.
About this article. This is for law firm innovation and IT leads, general counsel teams and product owners at legal-tech companies who are deciding whether to build on OpenAI's legal configuration. We checked every price, model ID, date and access rule on 20 November 2026 against OpenAI's pricing page, model page, API changelog, data controls and data residency guides, background mode guide and its two Astra for Law announcements (read through archived copies, as openai.com refused our plain requests), and checked the benchmark figures against the Artificial Analysis and Vals AI leaderboards. Rupee and sterling figures use European Central Bank reference rates of 19 November 2026, crossed and rounded to two decimals. This is not legal or financial advice, OpenAI can change prices and access terms, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What OpenAI actually lists
The pricing page groups the two legal IDs with other specialised models, such as the Daybreak cyber models and the Rosalind life sciences models, rather than with the flagship models. The table below shows the Standard rates per million tokens as published on 20 November 2026.
| Model ID | Context | Input | Output | Cached input | 30-minute cache write | 12-hour cache write |
|---|---|---|---|---|---|---|
| gpt-6.2-astra-law | Short | $8.00 | $40.00 | $0.40 | $10.00 | $20.00 |
| gpt-6.2-astra-law | Long | $16.00 | $60.00 | $0.80 | $20.00 | $40.00 |
| gpt-6-astra-law | Short | $10.00 | $50.00 | $1.00 | not listed | not listed |
| gpt-6-astra-law | Long | $20.00 | $75.00 | $2.00 | not listed | not listed |
| gpt-6.2-astra (base) | Short | $8.00 | $40.00 | $0.40 | $10.00 | $20.00 |
| gpt-6.2-astra (base) | Long | $16.00 | $60.00 | $0.80 | $20.00 | $40.00 |
Three things stand out. First, the legal configuration carries no premium today: every gpt-6.2-astra-law rate matches the base gpt-6.2-astra rates, and gpt-6-astra-law matches the short-context Standard input, output and cached rates of gpt-6-astra. Second, a single note under the specialised table says the Law and Rosalind promotional pricing is available at least through 31 December 2026. It does not say what the price becomes, or whether it changes at all, so treat any 2027 number as unknown. Third, the page shows the legal IDs only in the Standard table. They do not appear in the Batch or Flex tables, and the specialised Fast table lists no models.
When the long-context rate applies
OpenAI's GPT-6.2 Astra model page explains the long-context rule: for requests with more than 272,000 input tokens, input and cache rates double, and output costs 1.5 times the short-context rate, for the whole request. The legal rows follow the same pattern. The model accepts up to 922,000 input tokens inside a 1,050,000-token context window, with up to 128,000 output tokens. Legal work is where long prompts happen: a data room extract, a bundle of pleadings or a full share purchase agreement with schedules can cross 272,000 tokens easily, and then every token in that request is billed at the higher rate.
Discounts that do not obviously apply
Three cost levers that work on the base model are not documented for the legal IDs:
- Batch. The Batch table lists gpt-6.2-astra at half price ($4.00 and $20.00) but has no legal row.
- Background completion windows. OpenAI's background guide, changelog and pricing page give 30% off for a two-hour window and 40% off for an eight-hour window on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra and GPT-6 Luna. The legal configuration is not named.
- Ultrafast. Ultrafast on GPT-6.2 Astra costs three times Standard through 28 February 2027 and six times from 1 March 2027 (Pacific time). No legal Ultrafast rate is shown.
If an overnight review job is cost-sensitive, ask your OpenAI account team in writing whether these modes are available for gpt-6.2-astra-law before you design around them.
What the legal configuration is
Astra for Law is not a separate model. OpenAI's 17 November post calls it a model configuration of GPT-6.2 Astra: the base model plus an expanded legal search index and instructions for researching a question, weighing authorities and writing the answer. When OpenAI first introduced it on 17 September 2026, the configuration ran on GPT-6 Astra and was offered to selected firms through Trusted Access in ChatGPT and Codex, with the API identifier gpt-6-astra-law announced as coming later.
The search index is the main difference you pay for, even if you currently pay nothing extra. In September OpenAI said it covered US case law, statutes, regulations, court rules and administrative decisions across more than 230 million URLs, with case law drawn from the Free Law Project's CourtListener collection. By 17 November OpenAI said the index had grown by 34%, adding over 80 million URLs including material from the UK, Australia and Canada, to more than 310 million URLs.
The table below compares what changed between the two releases, according to OpenAI's own posts.
| Item | September 2026 release | November 2026 release |
|---|---|---|
| Base model | GPT-6 Astra | GPT-6.2 Astra |
| API model ID | gpt-6-astra-law (announced as coming) | gpt-6.2-astra-law (priced on the API page) |
| Who could get it | Selected law firms via Trusted Access | Any ChatGPT Enterprise customer can request it |
| Legal index size | More than 230 million URLs | More than 310 million URLs |
| Jurisdictions named | United States | United States, plus UK, Australia and Canada added |
| Named API builders | Harvey and Legora | Harvey quoted on results |
Who can use it, and how
Price is irrelevant if you cannot get access. OpenAI's 17 November post says every ChatGPT Enterprise customer can now request Astra for Law through the Trusted Access Program in the admin console, that teams with access select gpt-6.2-astra-law in ChatGPT, and that customers can request API access through a form. That means three practical paths:
- Law firms on ChatGPT Enterprise. An administrator requests access in the admin console. Lawyers then pick the legal configuration in ChatGPT, with no token billing to manage at matter level.
- Firms building their own tools. Request API access separately. Approval is not automatic, so plan your pilot timeline around the wait.
- Legal-tech vendors. In September OpenAI named Harvey and Legora as API customers building on the configuration. A smaller vendor should apply early and build its product so that the model ID is a setting, not a constant in code.
Data handling for client confidences
In September OpenAI said eligible firms in the programme get Zero Data Retention on the API, and that ChatGPT Enterprise usage is excluded from human review by default. OpenAI's data controls guide sets out what Zero Data Retention changes: customer content is kept out of abuse-monitoring logs, and the store parameter on the Responses and Chat Completions endpoints is always treated as false. The same guide notes that approval is required, that some endpoints and capabilities are not eligible and may still store application state, and that OpenAI reserves the right to make a model ineligible for a specific customer after written notice. We covered the endpoint gaps in our Zero Data Retention analysis. Ask for the ZDR approval for the legal configuration in writing, not just for your organisation in general.
What the benchmarks say, from three sources
OpenAI's headline claim is first place on Harvey LAB-AA, the Artificial Analysis version of Harvey's Legal Agent Benchmark, which uses 120 assignments across 24 areas of law and penalises material hallucinations. The Artificial Analysis leaderboard confirms the order: Astra for Law 6.2 at maximum reasoning effort scored 20.8%, ahead of GPT-6.2 Astra at maximum effort on 15.0% and at high effort on 13.3%. OpenAI also reports the best model from outside OpenAI, Meta's Muse Wonder Full, at 13.1%, which the leaderboard confirms.
Cost per task is where firms should look harder. Artificial Analysis records about $7.33 per assignment for Astra for Law 6.2 at maximum effort, against $6.35 for GPT-6.2 Astra at maximum effort and $3.86 at high effort. So the legal configuration scored about 39% more full-credit results than base GPT-6.2 Astra at maximum effort (20.8 divided by 15.0), at roughly 15% more cost per task. OpenAI says it averaged about 64,000 output tokens per assignment, against 139,000 for Muse Wonder Full.
| Benchmark | Astra for Law result | Comparison | Source |
|---|---|---|---|
| Harvey LAB-AA, full credit | 20.8% | GPT-6.2 Astra (max) 15.0% | Artificial Analysis |
| Harvey LAB-AA, cost per task | About $7.33 | GPT-6.2 Astra (max) about $6.35 | Artificial Analysis |
| Harvey LAB-AA, material hallucinations per task | 0.03 | GPT-6.2 Astra (max) 0.06, (medium) 0.03 | Artificial Analysis |
| Vals AI Legal Research Bench, all-pass | 53.85%, rank 7 | Leader Gemini 4 Neon 56.25% | Vals AI |
| Vals AI Legal Research Bench, cost per test | $2.75 | Gemini 4 Neon $7.37 | Vals AI |
The Vals AI leaderboard gives a more balanced picture than the launch post. Astra for Law sits seventh, level with Gemini 4 Hydrogen and 2.4 points behind the leader, but it is cheaper per test than five of the six models above it, and OpenAI says its pass rate there rose from about 46% in September to 53.8% on Vals AI's 16 November run. On hallucinations, Artificial Analysis lists Astra for Law at 0.03 material hallucinations per task. Base GPT-6.2 Astra recorded 0.06 at maximum effort but 0.03 at medium, high and extra-high effort, as did GPT-6.2 Sol at high effort and GPT-6 Astra at maximum effort. In other words, the low hallucination rate is largely a property of OpenAI's recent models, not something only the legal configuration provides.
None of these tests covers Indian law. The indexes and benchmarks are built around US law, and OpenAI names only the UK, Australia and Canada as added jurisdictions.
A worked cost model
Here is a simple model for one research or drafting assignment, at Standard rates, using OpenAI's own 64,000 output-token average.
- Short-context assignment with 150,000 input tokens: 0.15 million times $8.00 is $1.20 for input; 0.064 million times $40.00 is $2.56 for output; total $3.76.
- Long-context assignment with 400,000 input tokens: 0.4 million times $16.00 is $6.40; 0.064 million times $60.00 is $3.84; total $10.24, about 2.7 times the short-context cost.
- Short-context with caching, where 120,000 of the 150,000 input tokens are a reused prefix such as a firm style guide and a matter bundle: 30,000 fresh tokens cost $0.24, 120,000 cached tokens cost $0.048, output stays at $2.56, total about $2.85. The first request that writes the 120,000-token prefix pays $1.20 for a 30-minute cache write.
At 2,000 short-context assignments a month, the bill is 2,000 times $3.76, or $7,520. At ₹93.30 and £0.76 to the dollar (ECB rates of 19 November 2026, crossed and rounded), that is ₹7,01,616 or £5,715.20 a month, before taxes. If the same 2,000 assignments each cross 272,000 input tokens, the bill becomes $20,480, or ₹19,10,784. The single biggest cost control is keeping prompts under the long-context threshold, usually through retrieval rather than pasting whole bundles; our long context versus RAG comparison sets out the trade-off.
| Scenario | Per assignment | 2,000 a month (US$) | 2,000 a month (₹) |
|---|---|---|---|
| Short context, no caching | $3.76 | $7,520 | ₹7,01,616 |
| Short context, cached prefix | About $2.85 | About $5,696 | About ₹5,31,437 |
| Long context, no caching | $10.24 | $20,480 | ₹19,10,784 |
The cached row excludes the cache write charge, which depends on how often the prefix expires. These are illustrations, not quotes: reasoning effort, tool calls and retries change token counts, so measure the usage object on your own matters. Web search tool calls, if your workflow adds them, cost $10.00 per 1,000 calls plus search content tokens at model rates.
What it means for law firms
Pilot during the promotion, but do not anchor on it. There is no extra cost today for the legal configuration over base GPT-6.2 Astra, so the only question for a pilot is whether it gives better answers on your work. Run the same 30 to 50 real research questions through both, with lawyers grading the results blind, before 31 December.
Budget 2027 with a range. OpenAI's note says the promotion lasts at least through 31 December 2026. For context, the pricing page also shows OpenAI running other promotions with fixed end dates, such as GPT-5.6 Sol's promotional pricing ending on 30 November 2026. Build your 2027 budget on today's rate and on a higher scenario, and decide in advance at what price you would switch matters back to the base model.
Keep lawyers accountable for citations. A low hallucination rate is not zero. Every authority the model cites still needs to be checked against the source before it reaches a client or a court.
Mind the jurisdiction gap. If your practice is mainly Indian, English, Singaporean or Gulf law, the legal index adds less than it does for US federal and state questions. Test on your jurisdiction before paying any future premium.
What it means for legal-tech vendors
Treat the model ID as configuration. In two months OpenAI moved the configuration from gpt-6-astra-law to gpt-6.2-astra-law. Expect it to move again as new base models ship. Your product should switch model IDs by setting, with an evaluation suite that runs before each switch.
Price your product with headroom. If you sell seats or fixed-fee plans built on gpt-6.2-astra-law, your gross margin depends on a promotional rate with no published end price. Write a pass-through or repricing clause into 2027 contracts, or cap usage per seat.
Plan for no Batch discount. Bulk jobs such as contract review across a data room are where vendors normally use Batch at half price. Until OpenAI lists a legal Batch rate, those jobs cost full Standard rates on the legal ID, so consider routing bulk extraction to base GPT-6.2 Astra on Batch and reserving the legal configuration for research and drafting steps.
Know what you are differentiating against. Any ChatGPT Enterprise firm can now request the same configuration directly. A vendor's value has to come from workflow, firm data, integrations and review controls, not from model access alone. OpenAI also launched 26 partner-built plugins in September, connecting ChatGPT to tools such as Relativity and Clio.
Data residency and India-specific considerations
For Indian firms and vendors, two documented limits matter. First, OpenAI's data residency guide lists India among regions with regional storage, but in India, Australia and Japan only gpt-6-luna is listed for regional processing, on Standard pay-as-you-go; GPT-6.2 Astra is documented with US and EU data residency only. The legal IDs do not appear in the residency tables at all. Second, regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026, which would take the $8.00 and $40.00 base rates to $8.80 and $44.00 where residency applies. We explained the India storage-versus-processing gap in our regional processing article.
Client files often contain personal data, so Indian firms should map any use against the Digital Personal Data Protection Act 2023: record the purpose, minimise what is sent, and keep the processing agreement and any ZDR approval on file. Billing is in US dollars, so rupee costs move with the exchange rate as well as with token counts. And because the legal index centres on US law, Indian practice teams should judge the configuration on Indian research tasks before treating benchmark scores as relevant to them.
A six-step plan before 31 December
- Request Trusted Access in the ChatGPT Enterprise admin console, and file the API access form if you will build tools.
- Ask OpenAI in writing for ZDR status on the legal configuration, and whether Batch, background windows and residency apply to gpt-6.2-astra-law.
- Build a 30 to 50 question test set from closed matters, graded blind by lawyers, and run it on gpt-6.2-astra-law and gpt-6.2-astra.
- Measure tokens per assignment and the share of requests above 272,000 input tokens.
- Set a 2027 budget with a base case at today's rate and a higher case, plus a fallback rule.
- Put the model ID behind a setting and re-run the test set before every switch.
OpenAI's other trusted-access models follow a similar pattern of separate pricing and access rules; see our Daybreak cost analysis for comparison.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our legaltech app development team can build a blind evaluation set from your matters, measure token costs on gpt-6.2-astra-law against the base model, and put model choice behind a setting before the promotion ends. Talk to us at /contact-us/.
Last updated: 20 November 2026.