Legal AI Pricing in 2026: Every Published Price, and Every Vendor That Publishes None

We checked 20 legal AI vendors' own sites on 22 July 2026. Six publish a price; fourteen route you to a sales call. The full table, the hidden costs, and what each price actually buys.

Search "how much does legal AI cost" and you will find a lot of pages that pose the question and then decline to answer it. We read the ten highest-ranking articles for this cluster before writing this one. Seven of them contain no dollar figure anywhere. Two of the remaining three flag their own numbers as approximate. Two literally ask "how much do legal AI tools cost?" in their FAQ section and answer "it varies."

It does vary. That is not a reason to publish nothing. So here is the boring version: a table of what each vendor actually says on its own website, the date we read it, and nothing else. No estimates, no directory listings, no "industry sources say." If a number is not on this page, it is because the vendor did not put it on theirs.

Key facts

• Six of the twenty vendors we checked publish a price you can read without talking to a salesperson: HAQQ, Claude (Anthropic), Paxton, Genie AI, open-source Mike OS, and Lexzur's tier structure without figures. On a strict reading — a dollar amount rendered on the page — it is five.
• The published entry range is $0 to $499 per user per month. Mike OS is free software you self-host and pay your own model bill for; Paxton is the highest published self-serve entry price at $499 per user per month.
• Fourteen vendors route every price question to a demo or a call: Harvey, Legora, CoCounsel, Lexis+ AI, Spellbook, Luminance, Everlaw, Eve, Alexi, Ivo, Robin AI, Clio Duo, plus Lexzur's unrendered figures.
• Trials substitute for prices more often than not. Lexis+ AI offers two days, Spellbook seven, Alexi a guided request. A trial tells you whether you like the product. It does not tell you what a five-lawyer firm pays in year one.
• The licence is rarely the largest line. Seat minimums, annual prepay, implementation fees, the research database underneath, and the lawyer-hours spent verifying output all sit outside the sticker.
• Nobody in this market prices the verification time, which in our own experience is the biggest recurring cost of running any of these tools properly.

Legal AI pricing 2026: the master table

Every cell was read on the vendor's own website on 22 July 2026. "Not published" means we could not find a figure on the pages we could load, on that date. It does not mean the product is expensive, cheap, or bad. It means you will learn the price on a call.

HAQQ's annual prices are lower than the monthly figures in that table: eFirm Boutique works out to $20.83 per user per month billed annually ($250 per user per year) and Purple to $50 ($600 per user per year). Chat annual plans are $300, $1,000 and $3,000 per year for Starter, Pro and Business. Those come from our pricing page, and if this table and that page ever disagree, the pricing page is right and this page is stale.

Tools we did not check, and why they are not in the table

A price table is only as good as its checking discipline, so the vendors below are listed as unchecked rather than quietly omitted. Several of them do publish prices. We are not going to reproduce a figure we did not read on the day.

Why half the market will not tell you the price

There is a tempting version of this section where the vendors are hiding something. That version is wrong, and it is worth saying so plainly before the rest of the page.

Enterprise software with a sales-led motion prices per deal. A firm-wide rollout involves a security review, an onboarding programme, data migration, integrations with a document management system, and a negotiated contract term. The scope genuinely differs from one buyer to the next, and a single number on a web page would be wrong for most of them. Harvey, Legora, Luminance, Everlaw and Eve all run that motion, and it is a legitimate way to sell software.

There is also a second reason, less flattering but equally ordinary: a published price is a competitive disclosure and an anchor you have to defend. Vendors who publish give up flexibility. Vendors who do not, keep it.

The consequence for you is the same either way. You cannot comparison-shop most of this category. You can comparison-shop the six vendors in the top half of the table. For the other fourteen, the first honest number arrives after a call, which means the evaluation cost of a shortlist is measured in calendar weeks, not browser tabs. If that matters to you — and for a solo or a five-lawyer firm it usually decides the outcome — your real shortlist is smaller than the market looks.

One more thing worth naming: the absence of a price is not evidence of a high one. We have no idea what Harvey charges, and neither does any article that tells you it does. Estimates for Harvey in particular circulate widely and get reprinted as if the vendor had confirmed them. It has not. A guessed number inside a table that claims to be verifiable poisons every other row.

The hidden costs: what you actually pay beyond the licence

The sticker is the part of the bill that gets negotiated in public. These are the parts that get negotiated in the second call, and they routinely exceed it.

Implementation and onboarding fees

Anything sold through a demo can carry a one-off implementation charge: configuration, template loading, DMS integration, training sessions. None of the fourteen sales-gated vendors publishes whether it charges one, because none of them publishes anything. Ask for it as a line item in the first call, not the fifth. The question to use is blunt: what is the total year-one invoice, including everything that is not the recurring licence?

Seat minimums

A per-seat price is only a per-seat price above the floor. HAQQ eFirm publishes its floor: two seats per workspace. Most vendors publish neither a price nor a minimum, which means a $X-per-seat quote can arrive attached to a ten-seat commitment and quietly become a $10X purchase. Spellbook states outright that its pricing is structured around the number of team members on the licence, without saying what the structure is.

Annual prepay

Annual billing is where published prices bend. HAQQ's own annual discount is roughly 17% on eFirm Boutique ($25 monthly against $20.83 effective) and the same shape on Purple ($60 against $50). Paxton publishes both a monthly and an annual figure — $499 per user per month, or $2,999 per user per year, which is a materially larger annual discount. Where a vendor publishes no price at all, assume the quote you get is the annual-prepay one and ask for the monthly number to see the real spread.

The research database underneath

This is the largest hidden cost in the category and the one most often left out of comparisons. CoCounsel's value comes substantially from sitting on Westlaw. Lexis+ AI sits on the Lexis corpus. Those subscriptions are a separate, historically expensive purchase, and a firm that does not already hold one is not buying the same product a firm that does is buying. Comparing a research platform's AI line item against a standalone drafting tool is comparing two different purchases. Compare the bundle.

Usage, overage and storage

Credit and usage models move the cost with the work. HAQQ Chat is credit-based with published overage rates that fall as the plan rises (20% on Starter down to 35% on Enterprise as a share adjustment against the standard rate) and published storage per tier, from 100 MB on the free plan to 1 TB on Business. Self-hosted open source inverts the model entirely: Mike OS is free as software and not free as infrastructure, because you pay your own model API bill and it scales with usage rather than headcount.

Verification time

No vendor prices this and every buyer pays it. In our 300-task frontier benchmark, 24% of 3,000 graded answers cited or applied law that did not support the claim. That number is about frontier models on that task set, not about the twenty products in the table above, but the operational implication carries: a lawyer checks the output. Whatever that lawyer's hourly rate is, multiply it by the checking time and you have a line item larger than most of the licences on this page.

What a frontier model costs versus what a legal platform costs

The floor of this market is not a legal product at all. A general assistant subscription starts at $20 a month, and in our own 19-tool benchmark a plain Claude chat window outscored most dedicated legal AI products. That is the uncomfortable baseline every priced legal tool has to clear.

Underneath the subscription is the raw inference cost, which is smaller than almost anyone assumes. We built a cost harness across modelled matters and logged per-call token counts against verified 2026 model prices. A full motion-to-dismiss workflow — research, draft, cite-check, revise — came to $1.67 in model inference. A single contract review came to about $0.027. A 100-document discovery first pass came to $0.38. Those are first-draft inference costs, not a finished filing.

The same motion-to-dismiss workflow run entirely on the most capable model, with no routing, cost $4.55. Same input, same output, same rubric; 2.7x the bill, from one architectural decision about which model runs which stage. The full write-up is in the $1.67 versus $4.55 routing experiment, and the structural version of the argument is in our work on ontology-driven cost reduction.

The point is not that legal platforms are overpriced. It is that inference is not what you are buying. You are buying jurisdiction handling, verification, matter context, permissions, audit trails and someone to call. A vendor whose price is far above the inference floor should be able to name which of those it adds. If it cannot, the honest comparison is against a $20 chat subscription.

Price per seat versus price per outcome

Almost everything in the table is priced per seat, which is a convenient billing unit and a poor match for how legal work arrives. Legal work comes in matters, and matters do not distribute evenly across seats. A per-seat licence charges the partner who uses it twice a month the same as the associate who lives in it.

• Per seat — predictable, easy to budget, penalises occasional users. This is what HAQQ eFirm, Claude, Paxton and most of the market do.
• Per credit or per usage — tracks the work, harder to forecast, rewards intermittent use. HAQQ Chat is credit-based for this reason; the free tier ships 20 credits and Business ships 1,500.
• Per matter — the unit lawyers actually think in, and the one almost nobody sells. We have not found a vendor in this table publishing a per-matter price.
• Bring your own infrastructure — Mike OS. Cost scales with usage, and the operational burden scales with it too.

The practical test: take last quarter's actual volume, not your headcount, and price it under each model. Firms with a few heavy users and many light ones are systematically overcharged by per-seat pricing, which is precisely the shape of most small and boutique practices. We wrote about that mismatch in the small-firm guide.

What is inside each price

A price is meaningless without the scope attached to it, and this is the place where a vendor comparing itself to others is most likely to cheat. So, specifically, what the two HAQQ prices buy:

HAQQ Chat ($0 free / $30 / $100 / $300 per user per month) is the AI surface: research, drafting, document analysis, a credit allowance per tier, and the mobile app. It is not a system of record.

HAQQ eFirm ($25 or $60 per user per month, minimum two seats) is practice management: a matter is an eleven-tab workspace, with judge, court, court case number and opposing counsel as native fields rather than free-text notes. Billing is real double-entry accounting with client trust accounts and a balance sheet, not an invoice generator. KYC is sanctions screening with dataset versioning and true-match adjudication. eFirm runs on desktop and web; there is no eFirm mobile app, and the mobile surface is HAQQ Legal AI Chat.

That scope is why we are not going to tell you HAQQ is cheaper than CoCounsel or Harvey "like for like." It is not like for like. A research platform bundled with a century of case law is a different purchase from a practice-management system with an AI layer, and both are different from a $20 chat subscription. The prices in the table are comparable as prices. The products are not.

How to run the procurement

If you take one thing from this page, take the question list. It is short and it is the part vendors are least prepared for.

• What is the total year-one invoice? Licence, implementation, training, migration, integrations. One number.
• What is the seat minimum and the contract term? Ask before the demo, not after the pilot.
• What is the monthly price, not just the annual? The spread between them tells you how much flexibility you are selling.
• Which database subscriptions does this assume I already have? For anything research-led, this is the question that changes the answer.
• What happens when we exceed the plan? Overage rate, storage caps, per-document or per-page charges.
• What is the renewal uplift, contractually capped or not? A cheap year one with an uncapped year two is not a cheap purchase.
• Who verifies the output, and how long does it take? Ask the vendor to walk you through a real check on your own document during the pilot, and time it.
• What does exit look like? Data export format, notice period, and what happens to your matters if you leave.

A vendor that will not give you a number before a demo may still be the right vendor. But you should know that is the deal you are entering, and price the evaluation weeks accordingly.

Disclosure

HAQQ published this page, HAQQ sells legal AI, and HAQQ appears in the table. That is a conflict of interest and there is no version of this page where it is not.

What we can do about it is make the page checkable. Every competitor cell is either a figure from that vendor's own site with the date we read it, or the words "Not published." We have not estimated a single competitor price, we have not lifted any figure from a software directory, and where we could not load a page we said so rather than filling the gap. If you find a vendor price we have marked "Not published" that is in fact published, that is a correction we want, and it is the fastest way to make this page worse for us and better for you.

Changelog

Five of the ten highest-ranking pages in this cluster carry a last-updated date identical to their publication date. Pricing decays faster than almost any other content type, so this page keeps a visible revision log and the date at the top moves only when something below actually changes.

• 22 July 2026 — First publication. Checked 20 vendor sites for a published price, a billing unit, a seat minimum and a self-serve path; recorded exactly what each page showed on the day. Six vendors published a figure or a tier structure; fourteen did not. Added the unchecked-vendor table so omissions are visible rather than silent. Declined to publish circulating third-party estimates for Harvey and GC AI.

Three things trigger a revision: a vendor publishing a price where it previously published none, a published price changing, and a HAQQ price change on our own pricing page. We aim to re-run the full sweep quarterly. If you spot a stale cell before we do, tell us and we will date the fix.

Key takeaways

• Most of this market does not publish a price. Fourteen of twenty vendors we checked route every price question to a call.
• Your self-serve shortlist is small. If you refuse to sit through a demo to learn a number, it is HAQQ, Claude, Paxton, Genie AI and Mike OS.
• The licence is rarely the biggest line. Seat minimums, annual prepay, implementation fees, the research database underneath, and verification time all sit outside the sticker.
• Inference is cheap; the product around it is what you pay for. A full motion-to-dismiss workflow cost $1.67 in raw model calls. Ask any vendor charging far above that what it adds.
• Per-seat pricing systematically overcharges firms with a few heavy users. Price your actual last-quarter volume, not your headcount.

• HAQQ pricing — the source for every HAQQ figure on this page
• The 19 best legal AI tools in 2026, ranked by benchmark
• The $1.67 versus $4.55 motion-to-dismiss routing experiment
• Cutting legal AI cost with a legal ontology
• Best AI for legal work: 300 tasks, 3,000 graded answers
• Legal AI for small law firms
• Harvey review and alternatives
• Lexis+ AI review versus Westlaw