You are deciding which AI coding assistant your engineering organisation standardises on, and your finance team has noticed that last quarter's bill did not match last quarter's forecast. Here is what that decision actually turns on.
All three leading vendors changed how they charge during 2026. GitHub Copilot moved to usage-based billing on 1 June. Cursor introduced a new premium seat tier from 1 July. Anthropic repriced Claude Team seats in January. None of these were price rises in the headline sense, and all three changed what a seat costs you in practice.
What do AI coding assistant seats cost in 2026?
Entry-level team seats sit between US$19 and US$40 per user per month, and premium tiers run from US$96 to US$125. Every vendor now meters usage above an included allowance, so the seat price is a floor rather than a total. Figures below are vendor-published as at 17 August 2026.
Published seat prices, all USD, per user per month
--- GitHub Copilot Business: $19 monthly, including 1,900 AI credits per user.
--- GitHub Copilot Enterprise: $39 monthly, including 3,900 AI credits per user, requires GitHub Enterprise Cloud.
--- Cursor Teams Standard: $32 on annual billing, $40 on monthly.
--- Cursor Teams Premium: $96 on annual billing, $120 on monthly. Five times the included usage of Standard.
--- Claude Team Standard: $20 on annual billing, $25 on monthly. Minimum two members.
--- Claude Team Premium: $100 on annual billing, $125 on monthly. Five times Standard usage.
--- Claude Enterprise (self-serve): $20 per seat monthly billed annually, plus usage billed at API rates.
One correction worth making early, because it circulates widely and is wrong. Claude Code is included on the Claude Team Standard seat, not restricted to Premium. Anthropic's own support documentation lists Claude Code access under Standard seat features. Budget accordingly.
What changed in GitHub Copilot billing on 1 June 2026?
Copilot replaced premium request units with GitHub AI Credits and moved every plan to usage-based billing. Base seat prices did not change. What changed is that model consumption is now metered against a credit allowance denominated in dollars, and the automatic fallback to a cheaper model was removed.
The mechanics, from GitHub's own announcement and its billing documentation:
--- Credits are consumed on token usage at published per-model API rates, covering input, output and cached tokens.
--- Code completions and Next Edit suggestions remain included and are not credit-billed. For a team whose usage is mostly autocomplete, the change is close to neutral.
--- Overage is $0.01 per AI credit. Included credits from each licence contribute to a shared organisation-wide pool, so heavy users draw on quieter colleagues' allowances.
--- Copilot code review additionally consumes GitHub Actions minutes at standard Actions rates.
--- Data-resident and FedRAMP-compliant requests carry a 10% model multiplier increase. Hong Kong enterprises with residency requirements should model this in rather than treat it as a rounding error. Our explainer on what data residency actually means covers when that requirement is genuine.
One time-limited detail. GitHub gave existing Business and Enterprise customers promotional included usage of $30 and $70 per month respectively, for June, July and August 2026 only. That promotion expires this month. If your Copilot spend has looked comfortable since June, September is when you find out what it actually costs.
The removal of model fallback matters more than it sounds. Previously, exhausting your allowance downgraded you to a cheaper model. Now it draws credits, and then bills overage. The failure mode changed from degraded quality to an invoice. If you have not set budget controls at enterprise, cost-centre and user level, set them this week.
How much does Cursor Teams cost after the July 2026 repricing?
Cursor kept the Standard seat at $40 monthly and added a Premium seat at $120 monthly, or $96 on annual billing. The change took effect immediately for new customers and from 1 July 2026 for renewing customers. Cursor also split included usage into two separate pools per seat.
The two-pool structure is the part teams misread. Every Teams seat now receives a first-party models pool, covering Cursor's Auto and Composer models, and a separate third-party API pool for external models. Exhausting one does not draw on the other.
This has a practical consequence. A team that standardises on a third-party frontier model will burn the third-party pool while the first-party pool sits unused, and will conclude the seat is expensive. A team that lets Cursor route to its own models will find the same seat generous. Cursor states that the Premium first-party pool should cover a full month of heavy agent use for 99% of users.
Two limits on the published figures, stated plainly. Cursor does not publish a numeric size for either pool, only relative language. And it does not publish an on-demand overage rate; usage beyond the allowance is billed in arrears at rates not disclosed on the pricing page. Cursor Enterprise, which is the only tier offering organisation-wide pooled usage alongside SCIM and audit logs, is custom-priced with no public figure.
How much does Claude Code cost per seat?
Claude Code is included on the Claude Team Standard seat at $20 per seat per month billed annually, or $25 monthly, with a two-member minimum. The Premium seat at $100 annually or $125 monthly provides five times the usage. Limits are per member and are not pooled across the team.
The usage distinction is what drives the seat choice, not the feature list:
--- Standard: 1.25 times Pro usage per session, one weekly all-model cap. Defaults to Sonnet 5.
--- Premium: 6.25 times Pro usage per session, two weekly caps, one across all models and one for Sonnet specifically. Defaults to Opus 5.
The absence of pooling cuts both ways. A team of twenty where three engineers do most of the agentic work cannot redistribute allowance to them, which argues for a mixed deployment: Premium seats for the heavy users, Standard for everyone else. Anthropic permits mixing seat types within one team. Team plans cap at 150 seats, above which you move to Enterprise.
For workloads that exceed any seat tier, API rates apply. Sonnet 5 is $2 per million input tokens and $10 per million output. Opus 5 is $5 and $25. Haiku 4.5 is $1 and $5. Batch processing saves 50%. Anthropic changed Sonnet pricing earlier in 2026, so verify current rates before modelling a year forward.
What does this cost a 40-engineer Hong Kong team per year?
At list price on annual billing, forty seats cost roughly US$9,120 on Copilot Business, US$15,360 on Cursor Teams Standard, and US$9,600 on Claude Team Standard. Those are floors. Metered usage, premium seat upgrades and residency multipliers sit on top of every one of them.
Worked annual list-price examples, 40 seats, USD, before usage
--- Copilot Business at $19 monthly: $9,120. Copilot Enterprise at $39: $18,720.
--- Cursor Teams Standard at $32 annual: $15,360. All-Premium at $96: $46,080.
--- Claude Team Standard at $20 annual: $9,600. All-Premium at $100: $48,000.
--- A realistic mixed Claude deployment, 8 Premium and 32 Standard: $17,280.
At an indicative HK$7.8 to the US dollar, the Copilot Business floor is around HK$71,000 a year and the mixed Claude deployment around HK$135,000. Convert at your own treasury rate; none of these vendors bills in Hong Kong dollars, so you also carry currency exposure that a per-seat table does not show.
The number that should worry a budget holder is not any of the above. It is the gap between the floor and the actual invoice. Copilot overage at $0.01 per credit is unbounded by default. A single team running heavy agentic workflows can add materially to that $9,120 in a month, which is precisely why budget controls exist and why September's Copilot invoice will be the first honest one.
Which should you standardise on?
There is no single winner, and any comparison that produces one is selling something. The right answer depends on where your code lives, how your engineers actually work, and whether you can tolerate variable monthly cost. Four defensible verdicts follow.
If your organisation runs on GitHub Enterprise Cloud and usage is mostly autocomplete: Copilot Business at $19. Completions stay uncharged, the pooled credit allowance smooths individual variance, and the tooling is already inside your existing identity and audit perimeter.
If your engineers work agentically in an editor for most of the day: Cursor, and budget for Premium seats for the top third of your team. The two-pool structure rewards teams that let Cursor route to its first-party models and punishes teams that pin everything to an external frontier model.
If you want predictable cost and terminal-based agentic work: Claude Team, mixed Premium and Standard. Per-member limits with no pooling make the bill boringly predictable, which is worth real money to a department head who has to defend a forecast.
If you have strict data residency or procurement constraints: none of the self-serve tiers. Copilot's 10% residency multiplier, Cursor's custom-priced Enterprise tier and Claude Enterprise all require a procurement conversation. Start it before you pilot, not after.
A note on the comparison most teams actually make. These tools are not substitutes for each other in the way a seat table implies. Several organisations run Copilot for completions across the whole engineering group and a smaller number of Cursor or Claude Code seats for the engineers doing agentic work. That is more expensive per head and frequently cheaper per outcome.
What these prices do not tell you
Published seat prices answer roughly half the question. The following gaps are real, and any vendor or partner who glosses over them is not helping you build a defensible budget. We would rather state them than let you discover them at renewal.
--- Two of three vendors do not publish overage rates. Only GitHub publishes a figure, $0.01 per credit. Cursor and Anthropic bill usage beyond allowance at rates not stated on their pricing pages. You cannot model a worst-case month from public information alone.
--- Cursor's usage pools have no published numeric size. "Five times Standard" and "generous" are the only quantities offered. A pilot is the only way to size it.
--- Every figure here is USD list price. Enterprise agreements, volume discounts and multi-year commitments are negotiated and not public. Assume list price is your ceiling, not your cost.
--- Seat cost is not programme cost. Rollout, code review policy changes, security review of generated code, and the engineering time spent evaluating tools are usually larger than the licence line in year one. The same pattern we described in why most enterprise AI pilots fail applies here.
--- Prices moved three times in eight months. Anthropic repriced in January, GitHub restructured in June, Cursor in July. Any twelve-month forecast built on today's page is an estimate, not a commitment.
And a limitation on our own position. UD does not resell GitHub, Cursor or Anthropic licences, holds no partner margin on any of them, and takes no commission on your choice. We have no commercial reason to prefer one. If you are a fifteen-person team with one repository, you do not need an advisory conversation; run a two-week pilot on the vendor already inside your stack and read the invoice. Our guide to enterprise AI seat pricing covers the adjacent question of chat assistant seats, which is a different budget line and a different decision.
The next step
Do not sign an annual commitment on this before you have one real invoice. The single most useful thing you can do this month is run a four-week pilot with fifteen to twenty engineers on your actual codebase, with budget controls switched on, and measure the delta between the seat floor and the invoice. That number, not any published table, is your business case.
Then ask the question that per-seat pricing hides entirely: which of the work you are about to buy AI seats for is engineering work at all? A meaningful share of what mid-market Hong Kong organisations describe as an AI tooling decision is actually operational capacity dressed as a licensing question, and it is worth separating the two before you commit a budget line for twelve months.
Twenty-eight years of Hong Kong enterprise work has taught us that the most expensive technology decisions are rarely the ones with the biggest price tag. They are the ones made without a real invoice in hand. We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team, Hong Kong. All prices in this article were verified against vendor-published sources, github.blog, docs.github.com, cursor.com, claude.com and support.claude.com, on 17 August 2026. Prices change frequently; confirm current rates with the vendor before committing budget. This article is general information, not financial advice.
Where UD fits
If part of what you are trying to buy is capacity rather than developer tooling, UD Cloud AI Staff is the alternative worth pricing alongside per-seat licences. It does not replace a coding assistant for your engineers, and we will say so plainly. For everything else your team is currently paying per seat to do, we'll walk you through every step, from mapping the work, to a scoped pilot, to the cost comparison you can take to your CFO. Twenty-eight years of Hong Kong enterprise experience, alongside you the whole way.
