Claude Code cost for a team is the per-token bill for every request the CLI sends while developers work, and Anthropic's published average is about $13 per developer per active day. This guide turns that number into a monthly budget for a 25-person team, shows what a heavy-user tail adds, and lists ten controls with the exact settings behind each one.
How Claude Code is billed
As of October 2026, Anthropic's Claude Code costs page reports that across enterprise deployments the average is around $13 per developer per active day and $150–250 per developer per month, and that 90% of users stay below $30 per active day. Anthropic recommends a small pilot group to establish your own baseline before a wider rollout.
How that cost reaches you depends on how developers sign in. Pro and Max subscribers have usage included in the subscription. On Team and Enterprise plans, each member draws from a per-seat allowance that resets on a rolling five-hour window and a weekly window, shared with Claude chat; members can continue past it only if an admin turns on usage credits. Through the Claude Console, a cloud provider, or a gateway such as ATP Token, Claude Code is billed per token to the organization.
This article is about the per-token case, where nothing stops a session except a ceiling you set yourself.
Claude Code cost for a 25-developer team, worked out
Take 25 developers who each use Claude Code on 20 working days a month, at Anthropic's $13 average:
25 developers × 20 active days × $13 = $6,500 per month
That is $260 per developer, slightly above Anthropic's $150–250 monthly band, because 20 active days assumes everyone uses it every working day.
Averages hide the tail. Suppose five developers run long sessions or parallel instances and land at 2× the average ($26 per active day):
The last row assumes one average developer runs an agent team for five days. Anthropic says agent teams use "approximately 7x more tokens than standard sessions" when teammates run in plan mode, so 7 × $13 = $91 a day instead of $13. ATP Token bills in credits, where 1 credit = USD 0.01.
10 controls for Claude Code spend
1. Give coding agents their own project
Create a project only for Claude Code (and other coding agents), separate from the project your production services use. Every key in a project spends from that project's balance, so an agent that loops all night cannot drain the budget a customer-facing service depends on. The pattern is described in one project, one key.
2. Issue one key per developer inside it
Keys belong to exactly one project and inherit its allowed models and balance. With one key per person, the Usage page shows credits and tokens by key, which is per-developer spend without extra tooling. Keys start with atp- and the secret is shown once (managing API keys).
3. Treat the allocation as the monthly ceiling
Allocate the heavy-tail figure from the table, 780,000 credits, to the project. A project can only spend what it was allocated; once the balance is gone, requests return 402 until an admin allocates more. If you turn on project auto top-up, set its monthly cap too, or the ceiling stops being one. Setup steps are in team with budget caps.
4. Allowlist claude-sonnet-4-6; put Opus in a separate project
Enable only claude-sonnet-4-6 on the main coding project. At ATP list rates it is $3 input and $15 output per million tokens; claude-opus-4-8 is $5 and $25, so every token costs about 1.67× more. If someone switches to a model the project does not allow, the gateway rejects the call with 403 before it reaches any provider. Developers who need Opus for architecture work get a second project with a small allocation, such as 50,000 credits.
5. Pin ANTHROPIC_MODEL
Set ANTHROPIC_MODEL="claude-sonnet-4-6" in each developer's environment or in the env block of ~/.claude/settings.json. The --model flag overrides ANTHROPIC_MODEL for a session, so pinning sets the default and the allowlist in control 4 enforces it.
6. Make /usage and /clear a habit
The Session block of /usage shows token usage and a dollar estimate that Claude Code computes locally at list price. The ATP Usage page shows what the project was actually charged. Anthropic's advice for keeping it low:
- Run
/clearwhen switching to unrelated work. Stale context is re-sent with every message, and/clearcosts nothing. - Lower the effort level with
/effortfor simple tasks. Thinking tokens are billed as output tokens. - Keep
CLAUDE.mdshort and move specialised instructions into skills that load on demand.
7. Cap headless runs with --max-budget-usd and --max-turns
For CI jobs and scripts, Claude Code's CLI reference documents two print-mode flags. --max-budget-usd stops the run once its estimated spend, including subagents, reaches the amount. --max-turns exits after a fixed number of agentic turns.
claude -p --max-budget-usd 5.00 --max-turns 30 "fix the failing tests in src/auth"
This is a per-run stop on the client side. The project allocation remains the server-side ceiling for the job that retried a flaky test all night.
8. Keep agent teams opt-in
Agent teams are disabled by default and switched on with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1. At roughly 7× the tokens of a standard session in plan mode, one developer running an agent team spends about what seven developers spend in standard sessions. Allow it in a separate project with its own allocation, and ask teams to use Sonnet for teammates and shut them down when the task is done.
9. Review spend weekly from request logs
Every Monday, open the Usage page sorted by key and by model, then check Request logs for the same week. Each row shows time, scope, model, status, request ID, and input/output tokens. Request-log retention is 7 days, so a weekly review is the longest gap that still sees every request; to keep history, pull the rows with the request logs endpoint. Look for keys above 2× the team median, models outside the plan, and runs of 4xx or 5xx statuses.
10. Revoke keys on offboarding
When a contractor leaves, revoke their key from the API keys page. A revoked key stops working immediately and stays in the roster for audit, and because each developer has their own key, nobody else's session breaks.
For the economics behind controls 6 to 8, see what is the agent tax. For which model to allowlist for which job, see the best LLM for coding guide.
Setting up Claude Code on ATP Token
Step 1: Create the project and allocation
In the console, create a workspace and a project named for the team (for example claude-code-platform), enable claude-sonnet-4-6 as the only allowed model, and allocate the monthly credits.
Step 2: Issue a key per developer
Create one key per person from the project and send each secret through your password manager. It is shown only once.
Step 3: Point Claude Code at the gateway
The base URL has no /v1, because Claude Code appends /v1/messages itself. Clear ANTHROPIC_API_KEY so it does not take precedence.
export ANTHROPIC_BASE_URL="https://api.atptoken.ai"
export ANTHROPIC_AUTH_TOKEN="atp-..."
export ANTHROPIC_API_KEY=""
export ANTHROPIC_MODEL="claude-sonnet-4-6" # any model from GET /v1/models
claude
The same values can go under the env block of ~/.claude/settings.json.
Step 4: Verify
Run one prompt, then open the console. The call appears under Request logs and the credits appear on the Usage page. Full steps: run Claude Code on ATP.
