What Claude Code costs on API billing, October 2026
The three prices behind a Claude Code bill, what prompt caching does to them, and worked examples at Anthropic's list price and ours.
Three prices
A Claude Code bill on an API key comes from three rates: input, cached input and output, each per million tokens. Claude Code sends the whole conversation again on every request, so most of the tokens in a session are the same tokens, read back from cache. The cached rate decides most of the bill.
| Model | Input | Cached input | Output | List input | List cached | List output |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | $2.10 | $0.12 | $6.00 | $4.00 | $0.20 | $20.00 |
| Claude Sonnet 5.5 | $1.20 | $0.12 | $3.00 | $2.00 | $0.20 | $10.00 |
| Claude Haiku 5.5 | $0.06 | $0.0125 | $0.15 | $0.10 | $0.01 | $0.50 |
USD per million tokens. Ours from the live catalog; list is the price on the developer's pricing page, the read date on each model page.
Claude Opus 5.5 is $2.10 in and $6.00 out here, against $4.00 and $20.00 at Anthropic: 47% under list on input and 70% on output. The other rows read the same way.
One turn, worked
Take one turn late in a session: 80,000 tokens read from cache, 2,000 new input tokens and 1,500 output tokens. Cache writes are left out.
| Model | At list price | Here | Saving |
|---|---|---|---|
| Claude Opus 5.5 | $0.054 | $0.0228 | 57% |
| Claude Sonnet 5.5 | $0.035 | $0.0165 | 52% |
| Claude Haiku 5.5 | $0.0018 | $0.0013 | 23% |
80,000 cached + 2,000 input + 1,500 output tokens, each at its rate per million.
The arithmetic for Sonnet 5.5 at list: 80,000 × $0.20 + 2,000 × $2.00 + 1,500 × $10.00, all per million, is $0.035. The same turn here is $0.0165.
The same turn after the cache expires
On an API key the prompt cache lasts five minutes by default (Anthropic's Claude Code cost docs, read 2026-10-12). Step away for longer and the next request pays the input rate for the whole context: 82,000 input tokens plus 1,500 output.
| Model | At list price | Here |
|---|---|---|
| Claude Opus 5.5 | $0.358 | $0.1812 |
| Claude Sonnet 5.5 | $0.179 | $0.1029 |
| Claude Haiku 5.5 | $0.009 | $0.0051 |
82,000 input + 1,500 output tokens, nothing from cache.
That one turn costs about five times the cached one. A long break in the middle of a session costs one of these; a session that keeps moving pays the cached rate.
A day of use
Three sizes of day, each made of the turn above with every turn hitting the cache. The sizes are assumptions, not measurements; change the turn count and the table scales with it.
| Day | Tokens in a day | Claude Opus 5.5, list | Claude Opus 5.5, here | Claude Sonnet 5.5, list | Claude Sonnet 5.5, here | Claude Haiku 5.5, list | Claude Haiku 5.5, here |
|---|---|---|---|---|---|---|---|
| Light, 40 turns | 3,200,000 cached, 80,000 in, 60,000 out | $2.16 | $0.912 | $1.40 | $0.66 | $0.07 | $0.0538 |
| Steady, 150 turns | 12,000,000 cached, 300,000 in, 225,000 out | $8.10 | $3.42 | $5.25 | $2.475 | $0.2625 | $0.2018 |
| Heavy, 400 turns | 32,000,000 cached, 800,000 in, 600,000 out | $21.60 | $9.12 | $14.00 | $6.60 | $0.70 | $0.538 |
Each turn is 80,000 cached + 2,000 input + 1,500 output tokens. Cache writes and cold turns after a break are not included.
For scale, Anthropic's own figure for its deployments is about $13 per developer per active day at list price (the same docs). That is context, not a forecast for your sessions.
What /usage shows
Claude Code computes the session cost on your machine from its token counts at Anthropic's list price (the cost docs). It is an estimate at list, whatever host you point it at. The real bill is in the console, which records what each request was charged.
Thinking tokens
Thinking tokens bill as output and cannot be switched off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or the Fable models (the cost docs). A lower effort level produces fewer of them.
Setup
Two variables. The Claude Code guide has the settings.json form and the model variables.
export ANTHROPIC_BASE_URL="https://api.uplinkswitch.com/v1"
export ANTHROPIC_AUTH_TOKEN="sk-uplink-live-..."
claudePrompt caching works through the gateway; cached input bills at the cached rate in the first table.
Lowering the bill
/clearbetween tasks, so the next request does not carry the last task's context.- The right model per task. Haiku 5.5 for subagents and small edits.
- A shorter CLAUDE.md. It is in every request.
- A lower effort level where the task allows it.
- Keep sessions moving. A turn after the five-minute cache expiry pays the input rate for the whole context.
Questions
Why is Claude Code expensive? Every request carries the whole conversation. At 80,000 tokens of context, even the cached rate adds up over hundreds of turns, and any turn after a break pays the input rate for all of it.
What does a month cost? Multiply a row of the day table by your working days. Twenty steady days on Sonnet 5.5 at list is $105.00; here it is $49.50.
Where do I check my own numbers? Export your Anthropic usage and re-price it on the savings page. Every Claude model's price is on the Claude pricing hub.