What a coding agent rollout costs for 50 engineers

Seats for 50 engineers

Copilot's seat price includes the same value in AI credits, so a team that stays inside its pool pays only for seats, plus Actions minutes for code review. Claude Enterprise works the other way: the seat buys access, and every token is billed. The cheapest seat does not tell you which rollout is cheapest.

Claude Code, Cursor, Copilot and Codex compared

Usage can cost more than seats

Every tool bills usage beyond the seat allowance in tokens or credits if you allow it, and some bill it from the first request. Anthropic publishes the clearest figures we found, for Claude Code: around $13 per developer per active day and $150 to $250 per developer per month across enterprise deployments, with 90% of users below $30 per active day.

Model choice moves the number a lot. On Anthropic's list prices, Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, twice as much as Sonnet 5.5, while Haiku 5.5 starts at $0.10 and $0.50. Long sessions, agents working in parallel and automated reviews add to it: Anthropic notes that agent teams use about seven times more tokens than standard sessions when teammates run in plan mode.

For 50 engineers, Anthropic's range means $7500 to $12500 a month in usage. On seat plans, part of that is covered by the included allowance, and we found no comparable published figures for Cursor, Copilot or Codex. Treat it as a starting point, not a forecast: your mix of heavy and light users decides the real figure, and two weeks of measurement on one team tells you more than any benchmark.

One-off costs

These are mostly people's time, and they come before any saving. They do not shrink if you pick a cheaper tool.

• Instruction files, permissions and CI access for each repository the agents will work in.
• Single sign-on, user provisioning, admin policies and spend limits.
• A measured pilot with a baseline, so cost and quality can be compared before and after.
• Security and data protection review, and in Germany, involving the works council where there is one, because usage dashboards can show what individuals do.
• Training for engineers and for the people who review agent changes.

Costs teams forget

• Review time. Every change an agent writes still needs a person to read it, and that time usually comes from your most senior engineers.
• Rework. Changes that are rejected or reverted cost tokens and review time and ship nothing.
• Idle seats. Seats assigned to people who stopped using the tool still cost money, and on Claude Team and Copilot, removing a seat mid-cycle brings no immediate refund.
• Review bots. Claude's Code Review averaged $15 to $25 per review at launch, and Copilot code review uses AI credits and Actions minutes.
• Surcharges. Data residency adds 10% on Copilot and on eligible Cursor models, and US-only inference on Claude Enterprise is billed at 1.1 times API rates. Cursor adds $0.25 per million tokens for eligible third-party models on Teams and Enterprise.

A worked example

Here is one way the numbers can add up for 50 engineers on GitHub Copilot Business. The seat price and credit rate are GitHub's; the usage figures are assumptions for illustration, not measurements.

• Seats: 50 × $19 = $950 a month, which includes 95000 AI credits for the whole organisation.
• Usage, assumed: 30 engineers run agents daily and use $40 of credits each, and 20 mostly use completions, which are not metered, and use $5 each. That is $1300, or 130000 credits.
• Beyond the pool: 35000 credits at $0.01 each, so $350, if your budget allows it.
• Total: $950 + $350 = $1300 a month before tax.
• If the team merges 1200 changes a month, the tool costs about $1.08 per merged change.

Now add review. If each of those 1200 changes needs ten extra minutes of review, that is 200 hours a month, and at almost any engineer's hourly cost that is more than every line above. How fast your reviewers can read agent changes moves the total more than the choice of plan.

Budget per merged change

Seat prices are easy to compare and say little. Cost per merged change puts seats, tokens and review time against work that actually shipped, and it moves when you change models, instructions or review rules. Measure it per team from the first week of a pilot, next to a quality measure such as reverted changes, so a cheaper month does not hide worse code.

Our two-week cost review for one teamBudgeting for the people who review AI workAllocating LLM costs to teamsLLM cost optimisation