AI Coding Tools: What One Developer Really Costs

The seat price is the number on the budget sheet. The usage line is the number that moves. This post shows how AI coding tools bill, what one developer can cost in a month, and where to look when the total grows.

Three ways you pay

Every AI coding tool bills in one of three shapes. Some tools mix them.

A flat seat with an allowance

You pay per person. Each seat includes a fixed amount of use.

Claude Code on a Team or Enterprise plan works this way. Anthropic says each member’s use draws from a per-seat allowance that resets on a five-hour window and a weekly window. Usage inside that allowance is not metered in dollars.

GitHub Copilot works the same way at first. GitHub says each plan comes with an allowance of GitHub AI Credits.

A seat plus metered usage

The seat is the floor. Use past the allowance costs extra.

Anthropic calls this usage credits. On Teams and Enterprise plans, admins can set spend limits for the organization, a group, or one member.

GitHub moved Copilot to this model in 2026. Its announcement describes usage-based billing in AI credits. The Copilot plans page lists the seat prices. On 2026-10-04 it showed $19 per seat per month for Business and $39 for Enterprise. Check the page before you budget, because these numbers change.

Raw API tokens

There is no seat. Each token costs money. The total has no ceiling unless you set a limit.

Anthropic says Claude Code charges by API token consumption. When a team signs in through the Claude Console, the usage bills per token to the organization. An API key on any provider works the same way.

Where Cursor and Codex fit

Cursor bills as a seat plus metered usage. Its pricing page says every plan includes a set amount of model usage. Past that amount, on-demand usage is billed in arrears. On 2026-10-04, the page showed $40 per user per month for Teams.

Codex fits two shapes. OpenAI’s Codex pricing page says ChatGPT plans include Codex usage limits. Plus and Pro users can buy credits past the limit. With an API key, you pay for Codex at API prices. That is raw API tokens.

Why an agent session costs more than a chat

A chat sends a question and gets an answer. A coding agent works in steps. It reads a file, edits it, runs a test, and reads the result. Each step sends the whole conversation again.

Anthropic describes this in its cost guide. Claude Code sends your full conversation with every request. Each tool call adds another request. A long session carries a long history. Prompt caching makes the repeat reads cheaper, and it does not remove them.

The AI agent cost model on this site shows the effect step by step. The numbers below come from its coding-agent scenario.

A worked example

The scenario has 20 steps. It starts from a 1,500-token prompt. Each step sends 900 tokens more of file and tool output, and the agent writes 350 tokens per step. That is 219,000 input tokens and 7,000 output tokens for one task.

At the Claude Sonnet 5 list price of $2 per 1M input tokens and $10 per 1M output tokens, one task costs about $0.51. The prices are from our LLM pricing table, checked on 2026-09-27.

Now assume a developer starts 15 tasks a day for 21 working days. That is 315 tasks. At $0.51 each, the month costs about $160.

A heavy user might start three times as many tasks. That is 945 tasks and about $480.

These figures are assumptions. Your task length, model and cache use will move the result. Anthropic reports a figure of its own: an average of about $13 per developer per active day across enterprise deployments. Treat both numbers as a starting point for a pilot, as Anthropic’s page suggests.

Where the spend shows up

Each provider reports its own slice.

• Anthropic. The Claude Console usage page shows API spend. The Console Claude Code dashboard shows spend per member.
• GitHub. Copilot is a GitHub billing product. GitHub billing reports show usage and AI credit consumption.
• OpenAI. The API usage dashboard shows usage and exports it. The Usage and Costs API returns the same data in code.

CostMon reads these reports through its connectors. It shows daily spend next to your cloud bill. It does not split spend by developer, API key or repo. This is what each connector reports:

• The Anthropic connector groups cost by model.
• The OpenAI connector groups cost by model and line item. You can scope a connector to one project. Put your coding agents in their own OpenAI project and connect it. That spend then shows as its own line.
• The GitHub connector groups cost by day, product and SKU. It does not split by user or repo.

For per-developer detail, use the provider consoles above. Then set a cost alert on the AI line. A jump in daily spend reaches you before the invoice does.

Controls that work

1. Set a spend limit where the tool bills. Anthropic lets you set workspace spend limits. Claude Teams and Enterprise plans can set limits per organization, group or member. GitHub lets you set budgets and alerts. A budget on Copilot seats only sends alerts. A budget on Copilot AI credits can stop usage at the limit.
2. Use a smaller model for routine work. Anthropic says Sonnet handles most coding tasks and costs less than Opus. You can switch with /model.
3. Clear context between tasks. Anthropic says stale context wastes tokens on every later message. Use /clear when the work changes.
4. Lower the thinking effort on simple tasks. Thinking tokens bill as output. Our post on thinking tokens explains the cost.
5. Watch the daily total. A budget that you read once a month finds a problem late.

For a loop that went wrong at scale, read runaway bills. To price your own volumes, use the LLM cost calculator.

The seat price is easy to read. Look at the usage line each week.