AI Coding Assistant Cost at Enterprise Scale: Per Seat, Per Token, or Per Server?

Every major enterprise plan we checked is priced per user, and each pairs the seat with an allowance or a meter of some form. None of the vendors below publishes a per-server or unlimited-developer plan. That is the shape of the market, and it serves a great many teams well.

GitHub moved Copilot to usage-based billing on 2026-06-01: each Business or Enterprise seat includes a monthly pool of AI credits, consumed at published model rates, and plan prices did not change (checked 2026-09-18). Cursor prices Teams per user with included usage per seat and on-demand usage beyond it (checked 2026-09-18).

Anthropic prices Claude Team per seat with Claude Code included, and publishes Enterprise as a seat price plus usage at API rates. Tabnine prices its platform per user on an annual subscription, with model tokens metered separately when Tabnine supplies the LLM (both checked 2026-09-18).

Four things, and each is documented by the vendors themselves.

Seats. The count tracks hiring and churn. GitHub’s own billing page shows the asymmetry: adding Copilot licences mid-cycle grows the organisation’s credit pool immediately, while removing them does not shrink it until the next billing cycle (GitHub Docs, checked 2026-09-18).

Allowances that reset. GitHub’s included AI credits do not carry over between months. Cursor’s included usage is allocated per user, does not transfer between team members, and resets each billing cycle (Cursor Docs, checked 2026-09-18).

Usage past the allowance. GitHub enables additional usage by default for organisations and enterprises unless an administrator disables it. Cursor enables on-demand usage by default on Teams and bills it in arrears at public list API prices; Cursor documents a Token Rate of $0.25 per million tokens for third-party models, but billing details and controls vary by plan and configuration — check the current terms. Anthropic bills extra Team and Enterprise usage at standard API rates once an organisation owner enables it (all checked 2026-09-18).

Renewal. Seat count, allowance and price list are all revisited at term end. All three vendors publish spend controls — GitHub budgets set in USD, Cursor monthly team-wide limits, Anthropic organisation spend limits — so each item is a legitimate control. Together they are a forecasting job that grows with the organisation.

Every figure below is a published list price, read on the vendor’s official page on 2026-09-18; prices change, so follow the source before you budget. Tiers marked “not published” are sold by quote.

The table is a like-for-like reading of published terms, not a ranking. Which shape fits depends on headcount, hardware and where the code must stay.

Take a team of 30 developers and multiply each published seat price by 30 seats and 12 months. The figures are seats only and illustrative: they exclude usage beyond the included allowance, GitHub Enterprise Cloud, hardware, deployment services and any discount a vendor may offer.

Read the last row carefully. NeueCode’s figure is not a per-year seat total but a one-time licence per server plus assurance from year two, and it excludes the GPU hardware you buy and run; the per-seat figures exclude usage past the allowance. The two shapes do not sit on one line, which is why we publish no savings figure — the pricing model page states the model plainly: one fixed price per active server.

Seats are countable; tokens are not, until the month is over. BCG’s 2025 analysis of B2B software pricing describes pure usage-based pricing as offering customers maximum flexibility while being the least predictable, and cites an Andreessen Horowitz survey in which 36% of respondents considering outcome-based pricing worried about cost predictability (BCG, 2025-08-13; checked 2026-09-18).

Vendors know this, which is why each vendor above publishes spend controls. When inference runs on a vendor’s cloud, the vendor meters it and publishes the rate; Tabnine’s price list makes the same point from the other side — provider prices plus a 5% handling fee when Tabnine supplies the LLM, unlimited when the customer runs its own (tabnine.com/pricing, checked 2026-09-18).

When the model runs on hardware you own, the per-token line is replaced by a hardware line: GPU servers you buy, power and operate, budgeted rather than metered. It is a different line, not a missing one.

NeueCode 7 — Govern AI Autonomous Enterprise System — is one perpetual licence per active server, unlimited developers on it within the server’s sized capacity, and no per-token licence fees from NeueCode. The licence price is set only by the number and role of licensed servers — never by a server's size: each active physical or virtual server needs its own licence, bought once, and a passive disaster-recovery server is licensed at a reduced rate. Developer count, concurrent agents, repositories and model choice size the hardware — never the price.

Software assurance is included in year one and runs annually from year two as a fixed share of the licence, and a fixed-fee scoped pilot comes first. For a budget owner that means a licence line per active server that does not move with headcount, no usage-based line items on the licence, and a renewal stated as a share of the licence rather than a re-count of seats.

What it does not change: you still buy and run the GPU hardware, including its electricity; a server has a capacity, and enough concurrent load calls for another licensed server; deployment services are priced separately; and optional third-party model providers, off by default, are billed by that provider if you enable one. The per-server figure comes from the price sheet — list price shared on request — and the pricing model page sets out the terms in full.

Per-seat pricing fits small and mid-sized teams, teams without GPU hardware, and organisations that have standardised on a cloud IDE assistant and are content with its terms. The monthly number is small, the controls are good, and there is nothing to run.

A per-server licence fits large engineering organisations where seat counts move and usage is hard to forecast; regulated enterprises — government, banking, energy, healthcare, telecom and government contractors — that run or plan to run GPU hardware inside their perimeter; and programmes where the model, the prompts and the code must stay inside the network with evidence to show it. It also fits budget-governed teams that need a licence line item that does not move with headcount.

Many organisations will run both: a cloud assistant where cloud is approved, and a local (sovereign) deployment where it is not. The honest comparison is not a single number but a shape — countable seats plus a metered line, or countable servers plus the hardware to run them.

Take the published list price of the per-seat plan you would actually buy and multiply it by the seats you would actually grant, on the annual or monthly basis the vendor publishes. Add the included allowance per seat and ask your engineering leads whether agentic workloads will exceed it; if so, read the overage terms and decide whether you will cap or pay. Note when the term renews and what the vendor’s policy says about added and removed seats.

On the other side, ask for a per-server list price and a sizing conversation: how many active servers your concurrency needs, whether you want a passive DR server, what assurance costs from year two, and what deployment services and hardware will add. Put both on one page — seats plus meter beside servers plus hardware — and let your own headcount and usage decide.

We publish no savings figure because there is no honest general one. The pricing model page explains the model: one fixed price per active server, quoted on request.