What AI coding costs per developer, and how to measure your own

How much does AI coding cost? For Claude Code, Anthropic's own documentation gives a reference point: across enterprise deployments, the average is around $13 per developer per active day and $150 to $250 per developer per month, with 90% of users staying below $30 per active day.

Your number can sit far from that average. It depends on the model, the size of your codebase, how long your sessions run and whether you pay per token or through a plan. The reliable way to learn your own figure is to measure it, and UsageAtlas is a free way to do that across several coding tools at once.

What drives the cost

Four things move the bill more than anything else:

• Model choice. Bigger models cost more per token. Anthropic's docs suggest Sonnet for most coding work and saving Opus for harder reasoning.
• Context size. Coding agents resend the conversation with each request. A session left open all day keeps carrying its history, mostly as cache reads.
• Output and thinking. Output tokens are priced higher than input, and Anthropic bills thinking tokens as output.
• Parallel work. Subagents and agent teams each run their own context, so token use grows with the number running.

Plan or pay per token

How you pay changes what the question even means.

On a Claude Pro or Max plan, Anthropic notes that the session cost shown by /usage is not what you are billed, because the usage is part of the plan. That figure is still useful as a measure of how heavy your work is.

Find your own number

Averages help with budgeting a team, but one developer's month can look nothing like the mean. To get your real figure:

1. Install UsageAtlas, a free, open-source desktop app for Windows, macOS and Linux.
2. Let it read the local history from the tools you already use. Codex, Claude Code, Cursor, OpenCode and a few others can contribute cost estimates.
3. Check the history view after a normal week. You get estimated cost by day, by tool and by model, with the token mix behind it.

UsageAtlas computes these figures from tokens and known model prices, so they are estimates, not invoices. Where a model's price is unknown, the cost shows as unavailable rather than zero. Prompts and responses are not read, and no account is needed.

If your company already pays through one API account with a usage page, that page gives exact spend for that account. A tracker earns its place when you mix tools or plans and want one figure.

Ways to bring the number down

• Clear the session between unrelated tasks so old context stops riding along.
• Match the model to the job and drop to a smaller one for simple steps.
• Lower the effort or thinking level when deep reasoning is not needed.
• Watch which tool and model carried the expensive days, then change that habit first.

FAQ

Is AI coding cheaper on a subscription? For steady heavy use a plan can be, but it caps usage with windows. Measure a typical week before deciding.

Why did my cost jump in one session? Long context, cache misses after a break, and many subagents are common causes.

Is UsageAtlas free? Yes. It is free and open source.