How Much Does Claude Code Cost? Measured, Not Estimated

Almost every page answering this question quotes the same number: $150–250 per developer per month, sourced from Anthropic's own enterprise guidance. It gets republished because it is the only published figure available — not because anyone checked it.

Viberank can check it. Developers here submit real usage: token counts and API-equivalent costs computed by ccusage from their own local session logs. That makes this, as far as we know, the only independent distribution of what heavy AI coding actually costs.

The distribution

Monthly API-equivalent spend per developer, one row per person, across 1,096 developers:

The mean is $3,532 — well above the median, because the distribution has a long right tail. The top 10% of developers account for 55% of all spend tracked here. If you take one thing from this page, take that: “average” is the wrong statistic for AI coding cost, and any page quoting a single average is hiding the shape.

Against the published range

Share of developers burning more than each threshold:

Why the two numbers differ

Both are real; they measure different things, and saying so plainly matters more than the headline:

• Figures are API-equivalent cost computed by ccusage from local session logs, not amounts actually billed. Most developers on the board pay a flat subscription, so this measures what their usage would cost at API list prices.
• The population self-selects: these are developers who chose to measure their own usage and publish it, which skews heavier than the average Claude Code user.
• One row per developer (their highest-cost submission), normalised to a 30-day month. Submissions covering fewer than 7 days are excluded because they extrapolate to meaningless monthly figures.

The self-selection is the big one. People who install a usage tracker and publish the result are not a random sample of Claude Code users — they skew heavy, and that is exactly why the median here sits well above enterprise guidance. What the distribution does show honestly is the shape of heavy usage, which no other public source publishes at all.

What actually drives the number

1. Model choice. Opus costs materially more per token than Sonnet, which costs more than Haiku. Defaulting everything to Opus is the single biggest reason bills balloon.
2. Cache reads. About 95% of all tokens tracked here are prompt-cache reads, billed at a fraction of input price. This is why a developer can burn billions of tokens without a proportional bill — and why raw token counts badly overstate cost.
3. Context size. Bigger repos and longer prompts mean more input tokens every turn.
4. Session length. Long agentic runs re-send context repeatedly; frequency matters as much as size.

Measure your own

Don't estimate against someone else's median. ccusage reads Claude Code's local logs and computes your exact tokens and API-equivalent cost; one command puts it on the board so you can see which percentile above you actually land in:

npx viberank-cli

The subscription calculator takes the same number and tells you which Claude plan your usage justifies. If the answer stung, read how to cut your AI coding bill — model routing and caching alone can halve a heavy bill.

Citing these numbers

The figures on this page regenerate hourly from live submissions. If you want them programmatically, the data page documents a free, unauthenticated JSON endpoint at /api/stats. Attribution to viberank.app is all we ask.