Guide
AI coding tool costs per developer: Claude Code, Cursor, Copilot and the rest
Updated
If your engineers use several AI coding assistants, you pay for them in several ways: Claude seats, Cursor and Copilot licences, API keys, credits. Each vendor reports usage in its own units, in its own console. This guide shows how to collect it in one place, read it per tool, team and developer, and check whether the spend is paying off.
Which AI coding tools can you measure?
Token Meter reads usage from Claude Code, Claude Enterprise, Codex, Cursor, GitHub Copilot, Gemini CLI, Kiro, Cline, Continue and Windsurf, plus any other tool that emits the standard OpenTelemetry GenAI token metric. Devin is not supported yet. What each tool reports differs, so some give cost and tokens and others only lines or credits.
Only counts are read: sessions, prompts, tokens, lines, commits, credits, and accept or reject decisions. Prompts, code, diffs and conversation content are never stored.
How does each tool's usage data arrive?
Through whatever channel the tool supports: the vendor's analytics API, an OpenTelemetry export, the AIScale agent on each developer's machine, or the Token Meter Router for Cline and Continue. The table shows what each tool gives and what it doesn't.
| Tool | How it connects | What you get | Not available |
|---|---|---|---|
| Claude Code | Anthropic Admin key, or OpenTelemetry on any plan | Sessions, lines, commits, PRs, tokens and estimated cost by model, edit accepts and rejects, per person | The Admin API doesn't cover Claude Code run through Bedrock, Vertex or Foundry |
| Claude Enterprise | Analytics API key | Claude Code activity per person, claude.ai chat and Cowork activity, assigned seat counts | Needs an Enterprise plan; about a day behind |
| Codex | OpenTelemetry logs on any plan, or your OpenAI connection with a ChatGPT workspace ID | Sessions, prompts, tokens and estimated cost by model over OpenTelemetry; threads, prompts and credits by client from the usage API, per person | The usage API needs ChatGPT Enterprise; no lines or commits |
| Cursor | Team API key | Lines added by AI, accepts and rejects, prompts by agent, composer, chat and inline edit, Tab use. On Enterprise, exact per-commit AI lines. | No cost, no tokens, no sessions, no OpenTelemetry export |
| GitHub Copilot | GitHub organization and token; VS Code Copilot Chat can also send OpenTelemetry | Lines and prompts by feature, accepted suggestions and AI credits per person; tokens by model over OpenTelemetry | About two days behind; people named by GitHub login |
| Gemini CLI | OpenTelemetry through a Collector | Sessions, tokens by model, tool accept and reject | No lines, commits or cost |
| Kiro | Kiro Enterprise's daily OpenTelemetry export | Credits, messages and conversations per person and client | No tokens |
| Cline | Cline Enterprise's OpenTelemetry logs, or the Router | Tasks, prompts, tokens and cost by model | Over OpenTelemetry, people are anonymous unless a Collector adds an email |
| Continue | Continue's data destination, or the Router | Tokens by model and provider | People are anonymous IDs |
| Windsurf | Service key | Credits used per person per day | Needs Enterprise with individual-level analytics on; nothing else reported |
| Any other tool | OpenTelemetry | Tokens by model and person, shown as “Other tools” | No sessions, lines or commits |
Claude Code on any plan can send OpenTelemetry with a few lines in ~/.claude/settings.json, or in Claude Code's managed settings to roll it out to everyone:
{
"env": {
"CLAUDE_CODE_ENABLE_TELEMETRY": "1",
"OTEL_METRICS_EXPORTER": "otlp",
"OTEL_EXPORTER_OTLP_PROTOCOL": "http/protobuf",
"OTEL_EXPORTER_OTLP_ENDPOINT": "https://<your ingest address>/v1/otlp",
"OTEL_EXPORTER_OTLP_HEADERS": "Authorization=Bearer <your connector token>",
"OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE": "delta"
}
}Connecting two routes for one tool doesn't double count: when a vendor API and OpenTelemetry or AIScale describe the same tool, person and day, the vendor row wins, and between OpenTelemetry and AIScale, OpenTelemetry wins. The Router is the exception. A Cline or Continue call sent through the Router and also reported over OpenTelemetry counts twice, so pick one route per tool. Setup for every route is in Coding agents.
How do you see cost per tool, team and developer?
Usage lands on the Developers pages. Agent usage shows each tool and team over 7, 30 or 90 days. Team shows average cost per active user and spend per tool. Individual lists AI spend and engineering output per person, with no score or ranking.
Tools don't all bill in money, so spend is shown on one of three bases, never added together:
- Pay-as-you-go spend where the tool bills it.
- API-equivalent value otherwise: subscription use valued at API rates. It is not money spent.
- Credits where that is all the tool reports.
Cursor's feed has no cost or tokens, and Cursor usage-based spend is not collected yet. Sessions and prompts are shown separately too, because they are different units: Claude Code, Gemini CLI and Cline report sessions; Cursor, Copilot, Codex and Claude chat report prompts.
A figure counts under a person only when the tool reports a member's email, or a GitHub login mapped to a member. The Team page shows how many tool identities are not linked, per tool. To link more:
- Ask developers to run AIScale with their own key, so their usage carries their email.
- Map GitHub logins to members (needed for Copilot).
- Add
user.emailas an OpenTelemetry resource attribute for tools that report no user.
In the By team table, teams under 5 people are combined so no individual can be read from a row. Pay-as-you-go coding tool spend also appears as its own line in the AI Bill total, left out when the provider that bills it already has its invoice imported. See how to track AI spend across providers.
Are your Claude seats being used?
Seat management measures how much of each Claude seat's weekly allowance is used and grades every seat, from Achiever to Dormant. Anthropic publishes no token or dollar quota for a seat, so the figure is built from readings rather than a fixed limit.
Each seat's use is one of three kinds:
- Reported when Claude showed a percentage.
- Calibrated estimate when usage is divided by an allowance learned from readings.
- Peer estimate otherwise.
To set it up (admins and owners):
- Add a plan: Max / Pro, Team or Enterprise. Each person belongs to exactly one plan.
- Add seats one at a time, or import a CSV of
email,tier. People who use Claude but hold no seat appear under Not on any plan. - Set each tier's seat price per month and, if you know it, its weekly allowance.
- Add usage readings: the weekly percentage a seat holder sees in Claude under Settings, Usage. Readings above 50% are the most accurate; one below 30% is ignored.
Seats are graded over the last 4 complete weeks as Achiever, Steady, Laggard, Dormant or Unknown, with “at cap” and “downgrade?” flags. The values are API-equivalent dollars of Claude Code usage, not money spent, and chat is included only if a feed reports it. Before a renewal, the Dormant seats and those flagged “downgrade?” are the ones to look at.
For other tools, Your agents › Seats and licences shows seat utilization, idle seats and renewals, with a vendor admin connector for GitHub Copilot seats.
Is AI coding tool spend paying off?
You can't prove it from usage alone, but you can put the spend next to what teams ship. AI Impact joins coding tool cost with merged pull requests from GitHub, AI-written lines from AIScale and resolved tickets from Jira, at team level. Every figure is an association between spend and output, not proof the tools caused it.
The main figure is cost per AI-authored pull request: a team's pay-as-you-go AI tool spend divided by its AI-authored merged pull requests. Subscription usage has no per-token cost and is left out; a team with only subscription usage shows its API-equivalent value, marked “equiv.”
Illustrative numbers
A made-up month, to show the method, not an expected result.
- A team's pay-as-you-go AI tool spend for the month is $1,800.
- GitHub shows 60 AI-authored merged pull requests from that team.
- Cost per AI-authored pull request: $1,800 ÷ 60 = $30.
- Once the team has 8 weeks of history, the trend compares its first four weeks with its last four.
Know the limits before you act on it:
- A commit counts as AI-authored only when it carries a marker a tool leaves, such as a co-author trailer, or an agent account made it. Stripped trailers, squash merges and tab completion leave no mark, so AI-authored figures are a floor.
- For seat-billed tools such as Cursor or Copilot, AI spend per pull request is what the tokens would cost at API list rates, not what you paid.
- Planning, review and work done without AI tools are not in any of these figures.
- Team size, seniority and repository type are not controlled. Don't use these figures to rank teams or people.
- Commits and pull requests come from GitHub only. GitLab and Bitbucket are not supported.
Details: AI Impact. If the spend that worries you is API traffic rather than coding tools, see how to reduce LLM API costs.
Connect your first coding tool
Create a workspace, then connect one tool under Developers › Data sources: a vendor key, an OpenTelemetry token, or an AIScale key on your own machine. Vendor syncs run right after you save the connection, then hourly.