The model underneath legal AI costs $10 per million input tokens and $50 per million output, and it is the only layer of this category with a published, checkable rate card. OpenAI lists gpt-6-astra-law to the cent. Harvey publishes no price at all. Legora publishes no price at all. Spellbook publishes the shape of its pricing and withholds the amount. Thomson Reuters puts one figure on its own site — $551.85 a month, labelled “Starting Price”, on a page for firms of three attorneys or fewer.
So the arithmetic runs bottom-up, the only direction available. A research memo costs about $1.50 at model rates; one $551.85 seat-month buys roughly 368 of them. The markup above that is retrieval, verification, workflow and indemnity — usually worth paying. Here is how to tell when it is not.
Who publishes what
Checked 4 November 2026 against each vendor’s own site.
| Vendor | Publishes a price? | What you actually get told |
|---|---|---|
OpenAI (gpt-6-astra-law) | Yes, to the cent | $10 in / $1 cached / $50 out per million. No access gate published. |
| Thomson Reuters (CoCounsel) | Partly | ”$551.85/month” as a Starting Price on the solo/small-firm page. Not stated per seat. |
| Spellbook | Shape only | ”Pricing is structured around the number of team members on your license.” 7-day trial. No amount. |
| Harvey | No | /pricing is a 404. No pricing link in nav or footer. “Request a Demo.” |
| Legora | No | /pricing is a 404. “Book a demo.” |
Thomson Reuters’ configurator states what it needs before it will quote: “your sector, number of attorneys, jurisdiction, and preferred plan duration,” and adds that “Online pricing is available to new customers only.” The small-firm page also lists Westlaw Advantage with CoCounsel at $415.35/month and Westlaw Edge at $116.48/month, and routes anyone with more than ten attorneys to a sales rep.
Two things follow. First, the only vendor-published figure in the category comes with no seat denominator, so it cannot be compared against a per-lawyer number without assuming one. Second, any per-lawyer figure you have seen for Harvey or Legora is reported, not published — including the $1,000–$1,200 per lawyer per month on our own Harvey AI review, which is a third-party estimate and is labelled as one below.
What a legal matter costs at model rates
The rates are published. The token counts are modelled — stated here so you can substitute your own. Conversion at the conventional four characters per token, so 272,000 tokens is about 190,000 words.
| Task | Input tokens | Output | At $10/$50 | At $20/$75 (long) |
|---|---|---|---|---|
| Research memo on a discrete question (≈20 authorities) | 120,000 | 6,000 | $1.50 | n/a — under the threshold |
| Deposition transcript analysis (multi-day) | 320,000 | 10,000 | $3.70 | $7.15 |
| Contract set review (≈150 agreements) | 500,000 | 25,000 | $6.25 | $11.88 |
Re-querying the same authorities is where the arithmetic gets strange. At the $1.00 cached-input rate, the research memo falls from $1.50 to $0.42 — the input side drops 90% and output becomes the majority of the bill. A firm that works the same corpus all week is paying mostly for what the model writes, not what it reads.
One caveat on that: OpenAI publishes cache write prices for general Astra ($12.50 per million for 30-minute retention, $25.00 for 12-hour) and publishes none for the legal configuration. So the cheap read has an unpriced setup cost attached.
The 272K gap is a 2x error bar on the biggest line
Look at the two right-hand columns above. For the two long tasks, the published uncertainty is 1.9x.
General-purpose GPT-6 Astra carries two rate cards — $10 / $1 / $50 at or below 272,000 input tokens, $20 / $2 / $75 above it. Every GPT-6 model on the page carries that split, and it has been the dominant variable in OpenAI pricing since Astra launched. The legal configuration publishes one flat row: no long-context tier, no batch, no flex, no Fast, no Ultrafast.
That is a strange omission on this workload specifically. Legal documents are the longest-context material in professional work, and 272,000 tokens is a threshold a single matter clears without trying. The one workload where the long-context surcharge is least avoidable is the one with no published long-context price. Until OpenAI publishes a row, price a legal workload twice and carry the spread. The general-Astra threshold behaviour, and which models have no cliff at all, is worked through in what a long prompt actually costs on each vendor; our API cost calculator applies the 272K threshold to general Astra as the closest available proxy.
The missing batch row is a second, quieter cost. Batch takes 50% off input and output on every standard GPT-6 model. Overnight document review is the canonical batch job, and on the legal configuration the discount is simply not quoted.
What one seat-month buys at model cost
This is the comparison no vendor page performs, and the reason to do it is not to embarrass anyone. It is to size the layer you are buying.
| Licence | Published? | Research memos ($1.50) | Contract-set reviews ($11.88) |
|---|---|---|---|
| CoCounsel Legal, $551.85/mo | Published (starting price) | 368 | 46 |
| Harvey, $1,000/lawyer/mo | Reported | 667 | 84 |
| Harvey, $1,200/lawyer/mo | Reported | 800 | 101 |
At the cached rate the first row becomes roughly 1,314 research memos a month. A lawyer who produces 1,314 research memos in a month is not a lawyer.
That is the point of the exercise. For any realistic individual workload, the model cost inside a legal AI seat is a rounding error on the seat price. The money is buying the other layers, and the sensible question is whether those layers are good — not whether the seat is expensive relative to tokens.
What the markup buys, and why it is usually worth it
When OpenAI shipped Astra for Law in September, it kept the part that matters. The configuration pairs the model with a legal search index over more than 230 million URLs of US case law, statutes, regulations, court rules and administrative decisions, refreshed daily, plus domain instructions on handling authority and firm-level governance including information permissions and ethical walls.
OpenAI’s own evaluation: 54.0% overall correctness on 200 legal research questions, against 38.7% for the same model using ordinary web search. Both numbers come from one model, so the 40% relative improvement is entirely retrieval. Retrieval is what the independent vendors sell, and it is what you cannot assemble from an API key and a weekend.
So the decomposition is clean. The specialisation is free — OpenAI priced the legal configuration at exactly the general model’s rate, with no premium and no gate, while pricing its life-sciences sibling at a 50% premium behind an approval programme. The index, the citation verification, the authority sensitivity and the indemnity are what the markup covers. Nobody publishes a price for those separately, at any vendor, which means the markup is measurable in total and opaque in composition.
Who should build and who should buy
| If you are | Do this |
|---|---|
| A solo or small firm | Buy. At 50 research tasks a month the model cost is ~$75 against a $551.85 starting price, and that $477 gap does not build an authority index. |
| An in-house team doing high-volume document review | Price the API path properly. At 400 contract-set reviews a month the model bill is ~$4,750 against $20,000–$24,000 for 20 reported Harvey seats. Volume, not headcount, is what moves this. |
| Filing anything that cites authority | Buy the verification layer. A 54.0%-correctness retrieval system is not a thing to approximate with web search at 38.7%. |
| A US public-sector buyer | Your rate rose 10% on 1 November 2026 for FedRAMP endpoints. Get in writing whether that compounds with the 10% residency uplift before the next invoice. |
| An EU or UK firm | The legal section carries no residency footnote at all. Confirm eligibility and uplift before assuming the general +10% applies — see which AI APIs you can buy with EU data residency. |
| Already paying for a general assistant | Check what you have. A ChatGPT business seat is not a legal research tool, and the gap is the index, not the model. |
The pattern worth carrying away is the one visible across the whole rate card: the frontier labs have stopped charging for capability and started charging for the envelope around it — jurisdiction at 10%, authorisation at 10%, speed at 2x and 6x — while list prices for raw intelligence keep falling. In legal AI that inverts the usual advice. The model is the cheap, transparent, commoditised part. Every layer that makes it usable in practice is sold by quote. Current rates for both sit in our pricing tracker.
What changed
- 4 Nov 2026: First published.
gpt-6-astra-lawre-verified at $10 / $1 / $50 against OpenAI’s API pricing page, with no long-context, batch, flex, Fast or Ultrafast row and no access gate;gpt-6-rosalindconfirmed at $15 / $75 behind a trusted-access programme. Harvey and Legora confirmed to publish no price (both/pricingpaths 404); CoCounsel’s $551.85, $415.35 and $116.48 monthly starting prices and Spellbook’s 7-day trial read from vendor pages the same day.
Frequently asked questions
What does the legal model itself cost?
$10.00 per million input tokens, $1.00 per million cached input tokens and $50.00 per million output tokens. OpenAI lists the model configuration gpt-6-astra-law under a 'Legal model configurations' heading on its API pricing page, and 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 and no published access gate — which is the opposite of how OpenAI prices its life-sciences configuration, gpt-6-rosalind, at $15 in and $75 out behind a trusted-access approval programme. The legal rate is marked promotional: 'GPT-6 Astra for Law promotional pricing is available at least through December 31, 2026.' Read that as a floor rather than an expiry — it commits OpenAI neither to raising the price on that date nor to holding it. Rates verified 4 November 2026.
Does that mean Harvey and CoCounsel are overcharging?
No, and the arithmetic on this page is not an argument that they are. It establishes what the markup is, which is a different and more useful thing than a verdict on whether it is fair. What sits above the model is a retrieval index over case law and statutes, citation verification, authority sensitivity, firm-level governance such as ethical walls, workflow, and — the line no API carries — somebody contractually on the hook if the output is wrong. OpenAI's own evaluation puts Astra for Law at 54.0% overall correctness on 200 legal research questions against 38.7% for the same model using ordinary web search. That 40% relative gap is not a model gap, because both numbers come from the same model. It is a retrieval gap, and retrieval is precisely the layer the vendors sell. For a firm running fewer than a few hundred research tasks a month, paying for it is straightforwardly cheaper than building it.
Why can I not get a price out of Harvey or Legora?
Because neither publishes one. Checked 4 November 2026: harvey.ai/pricing returns a 404 and the site carries no pricing link in its navigation or footer at all — the only commercial route is 'Request a Demo'. legora.com/pricing also returns a 404, with 'Book a demo' as the sole call to action. Spellbook is a step more forthcoming: it publishes that 'pricing is structured around the number of team members on your license' and a 7-day free trial, but no amount. Thomson Reuters is the only vendor here that puts a dollar figure on its own site, and only on a landing page for firms of three attorneys or fewer. Everything else is behind a form that asks for your sector, attorney count, jurisdiction and plan duration first. Treat any per-lawyer figure you see quoted for Harvey or Legora as reported, not published — including the one on our own Harvey review.
What is the biggest unknown in the legal rate card?
There is no long-context price for it, on the most long-context workload in professional services. General-purpose GPT-6 Astra publishes two rate cards: $10 / $1 / $50 at or below 272,000 input tokens, and $20 / $2 / $75 above it. The legal configuration publishes a single flat row with no short/long split, no batch rate, no flex rate and no Fast or Ultrafast rate. Legal documents cross 272,000 tokens — roughly 190,000 words — routinely: a contract set, a document production, a multi-day deposition transcript. So a buyer modelling a legal workload cannot tell whether the dominant line of the invoice bills at $10 or $20 per million, which is close to a 2x error bar on the largest number in the model. The missing batch row costs money a second way, because batch takes 50% off input and output on every standard GPT-6 model and overnight document review is exactly the job batch exists for.
At what volume does building beat buying?
Later than most build-versus-buy instincts suggest, because the model is the cheap part. At 50 research memos a month the model cost is about $75 against Thomson Reuters' published $551.85 starting price — a $477 gap that will not buy you a 230-million-URL authority index, a citation verifier and an indemnity. At 400 contract-set reviews a month the model bill is roughly $4,750, against $20,000-$24,000 for a 20-seat licence at the figures reported for Harvey. That is where the arithmetic turns, and it turns on document volume rather than headcount. The honest threshold is not a number of seats but a question: do you already have the retrieval layer? If your authority corpus, citation checking and conflict walls are someone else's product, the API price is not the price of replacing them.
Do the compliance surcharges apply to the legal model?
Unstated, which is itself worth knowing before you budget. OpenAI charges 10% more on regional-processing endpoints for data residency, for 'models released on or after March 5, 2026', and from 1 November 2026 charges a further 10% uplift on FedRAMP endpoints. Both footnotes sit under the flagship model tables. Neither appears in the Legal model configurations section, so whether the legal configuration is eligible for data residency at all — and at what uplift — is not something the rate card answers. OpenAI also does not say whether the two 10% surcharges compound, so a public-sector buyer with a residency requirement cannot tell from the page whether they pay 10% or 21%. For a regulated firm those are the lines that grow, not the per-token rate.