OpenAI priced Astra for Law at exactly Astra's price — and from today charges 10% more for FedRAMP
When OpenAI launched Astra for Law in September 2026, the one number missing was the price. This desk said so at the time: no rate existed publicly, not for Trusted Access and not for the eventual API, and that single number would decide “whether this compresses the category or just adds a premium tier at the top of it.”
The number is now published, and it is the compressing answer.
$10 and $50, the same as the general model
A new section headed “Legal model configurations” has appeared on OpenAI’s API pricing page. It contains one entry:
| Model configuration | Input | Cached input | Output |
|---|---|---|---|
gpt-6-astra-law | $10.00 | $1.00 | $50.00 |
Per million tokens. Set that beside the general-purpose flagship on the same page:
| Model | Input | Cached input | Output |
|---|---|---|---|
gpt-6-astra (short context) | $10.00 | $1.00 | $50.00 |
Identical to the cent. OpenAI is charging nothing for the legal specialisation. The page also calls it a “model configuration” rather than a model, which is the same word it uses for gpt-5.6-cyber — a label for a tuned deployment of an existing flagship rather than new weights.
The rate is marked promotional: “GPT-6 Astra for Law promotional pricing is available at least through December 31, 2026.” Read that phrasing carefully, because it is a floor and not an expiry. OpenAI uses the identical construction elsewhere on the page for GPT-5.6 Sol, whose promotional pricing runs “at least through November 21, 2026.” Neither is a commitment to raise prices on that date, and neither is a commitment not to.
The contrast that gives the number meaning
OpenAI has shipped exactly one other vertical configuration, and it is priced and gated completely differently.
gpt-6-astra-law | gpt-6-rosalind | gpt-6-astra | |
|---|---|---|---|
| Input / output per M | $10 / $50 | $15 / $75 | $10 / $50 |
| Premium over Astra | None | +50% | — |
| Access gate | None published | Trusted-access approval program | None |
| Long-context rate | Not published | $30 / $112.50 | $20 / $75 |
Rosalind, the life-sciences model, costs half again as much as the flagship it is built on and requires approval to call. The legal configuration costs the same as the flagship and, as published, requires nothing. One vertical is sold as a controlled good; the other is sold as a list-price commodity.
The obvious reading is risk rather than economics. A biosecurity model carries a dual-use problem that a case-law model does not, and the premium is plausibly the cost of the gate rather than the cost of the capability. But it leaves OpenAI’s price list making a clear statement: domain specialisation, where it is not dangerous, is free.
What OpenAI does charge extra for, as of today
The things that carry a surcharge on this page are not capabilities. They are wrappers — and a second one started billing today.
A footnote under the flagship tables now reads: “Starting Nov 1, 2026, FedRAMP endpoints are also charged a 10% uplift.”
The word “also” matters. OpenAI already charges 10% on regional-processing endpoints — the data-residency route — for “models released on or after March 5, 2026.” FedRAMP endpoints, the authorisation path for US federal workloads, were outside that surcharge and from today are inside it. There is no announcement, no changelog entry and no notice period: the effective date is the day the footnote describes.
So the complete list of things OpenAI prices above standard rates is now: where you run it (regional processing, +10%), under whose authorisation you run it (FedRAMP, +10%, from today), and how fast you run it (Fast 2x, Ultrafast 6x). Not what the model knows.
One question the page does not answer, and a federal buyer needs it answered: whether the two 10% uplifts compound. An endpoint that is both FedRAMP-authorised and regionally pinned either costs 10% more or 21% more, and the rate card is silent. If you are budgeting a public-sector deployment, get that in writing before the next invoice rather than after it. The same ambiguity exists at Mistral, whose docs price a regional uplift and a priority tier separately without saying whether they stack.
The gap in the legal rate card
The most consequential thing about gpt-6-astra-law is a row that is not there.
General-purpose Astra publishes two rate cards: $10 / $1 / $50 below 272,000 input tokens, and $20 / $2 / $75 above it. Every GPT-6 model on the page carries that long-context second tier, and it has been the dominant variable in OpenAI pricing since Astra launched with the 272K cliff in September.
The legal configuration publishes a single row. No long-context rate. No batch rate, no flex rate, no Fast or Ultrafast rate.
Consider what a legal workload looks like. A contract set, a brief with exhibits, a deposition transcript, a document production — these are the longest-context documents in professional work, and 272,000 tokens is roughly 200,000 words, which a single matter clears without difficulty. The one workload where the long-context surcharge is least avoidable is the one workload with no published long-context price. A buyer modelling legal research cannot currently tell whether the dominant line of their invoice bills at $10 or $20 per million input tokens, which is a 2x error bar on the largest number in the model.
The missing batch row costs money in a second way. Batch processing takes 50% off input and output on every standard GPT-6 model; on the legal configuration it is simply not quoted. Overnight document review is exactly the kind of job batch exists for.
What to do about it
If you buy legal AI software, the number above is a floor on nobody’s costs but OpenAI’s, and that is its value. Whatever a legal AI vendor charges over $10 in and $50 out is not model cost. That arithmetic is now worked out per matter — and set against the one vendor seat price anybody publishes — in what legal AI costs above the model. It is the retrieval index, the citation verification, the workflow and the indemnity — which is precisely the argument this desk made in September, when the plugin partner list turned out to contain the two vendors whose entire product was “GPT, for law.” The specialisation was never the moat, and OpenAI has now confirmed it by pricing it at zero.
If you are building on the API, do not model a legal workload on the single published row. Price it twice, once at $10 and once at $20 per million input, and treat the spread as the uncertainty you are carrying until OpenAI publishes a long-context rate. Our API cost calculator applies the 272K threshold for the general Astra model, which is the closest available proxy.
If you are a FedRAMP customer, your rate went up 10% today, on no notice, and you should confirm with OpenAI whether that stacks with regional processing before assuming it does not.
The wider pattern is worth naming. Across 2026 the frontier labs have converged on charging for the compliance envelope rather than the intelligence inside it — residency at 10% nearly everywhere and 20% on Microsoft’s EU Data Zone, speed tiers at 2x and 6x, authorisation now at 10% — while list prices for raw capability keep falling, as GPT-6.1 Terra’s $1 rate showed four days ago. For a regulated buyer that inverts the usual advice. The model is the cheap part. The permission to run it where your lawyers require is the line item that grows.
Frequently asked questions
What does GPT-6 Astra for Law cost?
$10.00 per million input tokens, $1.00 per million cached input tokens and $50.00 per million output tokens, listed under a new 'Legal model configurations' heading on OpenAI's API pricing page as the model configuration gpt-6-astra-law. Those three figures are identical to the cent to general-purpose GPT-6 Astra at short context. There is no premium for the legal specialisation. OpenAI marks the rate promotional — 'GPT-6 Astra for Law promotional pricing is available at least through December 31, 2026' — which is a floor rather than an expiry date, so the rate may hold past that or may rise with whatever notice OpenAI chooses to give.
Is there an approval process to use it?
None is published, and that is the sharper contrast. OpenAI's other vertical configuration, the life-sciences model gpt-6-rosalind, costs $15 in and $75 out — a 50% premium over Astra — and sits behind a trusted-access approval program. The legal configuration carries no access note, no eligibility gate and no approval program anywhere on the pricing page: it appears as an ordinary model ID with an ordinary rate. Treat that as what the page says rather than as a guarantee, because OpenAI publishes no launch date for the configuration either, and a model ID can be listed before it is generally callable.
What changed for FedRAMP customers on 1 November 2026?
Bills rise 10%. A footnote under OpenAI's flagship model tables now reads: 'Starting Nov 1, 2026, FedRAMP endpoints are also charged a 10% uplift.' That is the same surcharge OpenAI already applies to regional-processing endpoints for data residency, which it charges on 'models released on or after March 5, 2026.' The word doing the work is 'also' — FedRAMP endpoints were not previously carrying the uplift, and from today they are. OpenAI does not say on this page whether the two surcharges compound for an endpoint that is both FedRAMP-authorised and regional, so a US public-sector buyer with a residency requirement cannot tell from the rate card whether they now pay 10% or 21%.
Does this mean legal AI vendors are overcharging?
It means whatever they charge above $10 in and $50 out is not model cost, which is a more useful statement than a verdict. This desk argued in September that the model was never the moat in legal AI — the proprietary search index over case law, statutes and regulations was, along with authority sensitivity, workflow and who carries the liability. The published rate is consistent with that: OpenAI gave away the specialisation and kept the retrieval layer. So when you evaluate a legal AI subscription, the question to ask is what the markup buys in retrieval quality, citation verification and indemnity, because it is demonstrably not buying a more expensive model.
What is the biggest gap in the published rate?
There is no long-context price for it, on the most long-context workload in professional services. General-purpose GPT-6 Astra publishes a second rate card above 272,000 input tokens — $20 in, $2 cached, $75 out, double the input price and 1.5x the output. The legal configuration publishes only a single short-context row, with no long-context, batch, flex, fast or ultrafast rate at all. Legal research means long documents: a brief, a contract set or a deposition transcript crosses 272,000 tokens routinely. Until OpenAI publishes a long-context row, a buyer modelling a legal workload cannot tell whether it bills at $10 or $20 per million input tokens above the threshold, and that is a 2x error bar on the dominant line of the invoice. The absence of a batch row removes the 50% discount from the model too.
Sources
- OpenAI API Docs — Pricing (primary: new 'Legal model configurations' section listing gpt-6-astra-law at $10.00 input / $1.00 cached input / $50.00 output per 1M; 'GPT-6 Astra for Law promotional pricing is available at least through December 31, 2026'; 'Starting Nov 1, 2026, FedRAMP endpoints are also charged a 10% uplift'; gpt-6-astra $10/$1/$50 short context and $20/$2/$75 long context; gpt-6-rosalind $15/$1.50/$75)
- OpenAI API Docs — Deprecations (primary: checked 1 November 2026; no GPT-6 family model deprecated, newest announcement 14 October 2026, no 31 December 2026 or 1 November 2026 shutdown date anywhere)
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.