What a real Claude Code session costs
Per-token price tables are easy to game: headline input and output prices get the discount treatment while cache pricing hides in a tooltip. But a real coding session is ~88% cache reads — the system prompt, project context, and conversation history hit the cache on every request. Cache pricing — reads plus writes — comes to about 56% of the session's bill at official rates, so it moves your cost more than the headline rate does.
The workload
A focused two-hour Claude Code session on a mid-size repository — a few dozen requests with the usual context reuse. The assumptions, stated plainly:
- Cache reads5M tokenscontext re-sent from cache on every request
- Fresh input250K tokensnew instructions, file contents, tool results
- Output100K tokensgenerated code and explanations
- Cache writes350K tokensnew context written into the cache
Your mix will differ — scale the numbers; the shape (cache-dominated) holds for any agentic coding tool.
The session, priced
Live prices from the marketplace — the same catalog that bills your requests. Official column uses the model owner's published direct-API prices.
| Model | Official API | KoRouter | You keep |
|---|---|---|---|
| claude-fable-5 | $10.97 | $5.91 / session | |
| claude-sonnet-5 | $2.19 | $1.18 / session | |
| gpt-5.6-sol | $1.35 | $5.40 / session |
Where the money goes — claude-fable-5
Cache reads $3.25 · fresh input $1.63 · output $3.25 · cache writes $2.84. Cache reads and writes are about 56% of this bill — which is why a provider that discounts headline prices but not cache pricing costs more than its badge suggests.
How to compare any two providers
- 1. Ignore the discount badge. Collect four unit prices from each: cache read, input, output, cache write.
- 2. Weight them by a realistic mix — for agentic coding, cache reads are 80–90% of tokens.
- 3. Compare the totals. A provider can undercut on output and still lose the session on cache.
KoRouter applies one multiplier to every price component of a model — cache included — so the discount you see is the discount your workload gets. Prices live on the marketplace; per-tool setup lives in Integrations.
Free to sign up, no monthly fee — pay as you go with credits.