Short answer
What one Claude Code session costs on an API key, at Anthropic's list prices, from real sessions between October 6 and November 2, 2026:
| Model | Median session | P75 | P90 | Median per hour |
|---|---|---|---|---|
| Claude Opus 5.5 | $6.96 | $18.82 | $44.17 | $10.93 |
| Claude Sonnet 5.5 | $2.22 | $6.53 | $14.96 | $4.89 |
| Claude Haiku 5.5 | $0.56 | $2.00 | $6.10 | $1.38 |
| Claude Fable 5.1 | $13.38 | $39.94 | $91.91 | $27.71 |
A session here is a run of Claude Code requests on one API key with no gap longer than 30 minutes. Half of Opus 5.5 sessions cost under $7 and nine in ten under $45. The costs are recalculated per request from Anthropic's pricing page, checked on November 4, 2026. About half of a typical session's cost is cache writes, not output.
Get an API key: top up from $5 by card or USDT and run Claude Code on Opus, Sonnet and Haiku, billed per token.
New accounts start with a $0 balance — top up before your first request. Top up from $5. If a request returns 402 (400 on the OpenAI format) with "wallet balance is insufficient", your balance is empty — your base URL and key are fine. Top up and retry.
Claude Code session costs by model
The same sessions, with how long they ran and how many model calls they made. Each model has at least 400 sessions in the sample.
| Model | Cost (median / P75 / P90) | Model calls (median / P75 / P90) | Minutes (median / P75 / P90) |
|---|---|---|---|
| Opus 5.5 | $6.96 / $18.82 / $44.17 | 42 / 127 / 311 | 39 / 93 / 181 |
| Sonnet 5.5 | $2.22 / $6.53 / $14.96 | 33 / 90 / 209 | 28 / 71 / 140 |
| Haiku 5.5 | $0.56 / $2.00 / $6.10 | 32 / 114 / 329 | 28 / 72 / 155 |
| Fable 5.1 | $13.38 / $39.94 / $91.91 | 32 / 78 / 229 | 28 / 79 / 148 |
How the sessions spread out:
| Model | Under $5 | Under $20 | Under $50 |
|---|---|---|---|
| Opus 5.5 | 41% | 77% | 91% |
| Sonnet 5.5 | 69% | 92% | 98% |
| Haiku 5.5 | 88% | 99% | 100% |
| Fable 5.1 | 23% | 62% | 80% |
The spread is wide. A quick question or a small fix costs cents, while one Opus 5.5 session in ten passes $44. Budget from the P75 or P90 column, not the median.
Cost per hour and per model call
For sessions of at least 10 minutes:
| Model | Per hour (median / P75 / P90) | Per model call (median) |
|---|---|---|
| Opus 5.5 | $10.93 / $18.09 / $28.48 | $0.14 |
| Sonnet 5.5 | $4.89 / $8.25 / $14.60 | $0.06 |
| Haiku 5.5 | $1.38 / $2.78 / $4.35 | $0.01 |
| Fable 5.1 | $27.71 / $41.51 / $71.21 | $0.40 |
A model call is one API request. One prompt you type usually triggers several calls, because Claude Code reads files, runs tools and sends the results back.
Why real sessions cost more than token estimates
Every call resends the whole conversation. In the median Opus 5.5 session, each call carried about 154,000 tokens of context (input plus cached input). On the three days we checked per call, about 8 in 10 Claude Code calls on Opus 5.5 carried more than 100,000 tokens.
Cache writes are about half the bill. In a median session, the cost split was:
| Model | Cache writes | Cache reads | Output | Uncached input |
|---|---|---|---|---|
| Opus 5.5 | 52% | 19% | 19% | under 1% |
| Sonnet 5.5 | 56% | 19% | 15% | under 1% |
| Haiku 5.5 | 43% | 28% | 17% | under 1% |
| Fable 5.1 | 64% | 9% | 22% | under 1% |
(Medians of each share, so a row doesn't add up to exactly 100%.)
A 5-minute cache write costs 1.25× the input price, while a cache read costs 0.05× on Opus 5.5 and Sonnet 5.5 and 0.025× on Fable 5.1. New context that lands in the conversation (file contents, tool output, your next prompt) gets written to the cache once, then read cheaply on later calls.
Caching already does most of the saving. About 90% of input tokens in a median session were cache reads. Billed as plain input instead, the same sessions would cost about 6 times as much on Opus 5.5 and Sonnet 5.5, and about 5 times on Haiku 5.5.
Haiku 5.5 often pays its long-context price. Haiku 5.5 costs $0.10 / $0.50 per million tokens only for prompts up to 100,000 tokens; above that, each request is billed at $0.50 / $2.50. In Claude Code, 61–71% of Haiku 5.5 calls were over 100,000 tokens on the days we checked, so most of them were billed at the higher rate.
How to keep a Claude Code bill down
These follow from where the money goes:
- Use Sonnet 5.5 for routine work. A median Sonnet 5.5 session costs about a third of a median Opus 5.5 session. Switch to Opus 5.5 or Fable 5.1 for the hard parts.
- Don't switch models mid-session without a reason. Caches belong to one model, so the first call on the new model writes the whole context again.
- Keep context short. Use
/clearwhen you start an unrelated task and/compactwhen a long session drags. Every call carries the whole context, so a smaller context makes every later call cheaper. - Mind long pauses. Anthropic's default prompt cache lasts 5 minutes. After a break longer than the cache lifetime, the next call writes the whole context to the cache again at the write price.
- Set a budget from P90. On Opus 5.5, $45 covers nine in ten sessions; on Sonnet 5.5, $15 does.
Running Claude Code through TeamoRouter
TeamoRouter bills the same Claude models per token, below Anthropic's list price (current prices), with no subscription. Point Claude Code at it with two environment variables:
export ANTHROPIC_BASE_URL=https://api.teamorouter.com
export ANTHROPIC_AUTH_TOKEN=your-teamorouter-key
claude
Full setup: Install Claude Code. Top up from $5 by card or USDT (how).
How we measured
- Sample: Claude Code requests on API keys routed through TeamoRouter from October 6 to November 2, 2026 (Asia/Shanghai days), identified by Claude Code's user agent (the terminal app and the IDE extensions). Only successful requests count.
- Session: requests on one API key with no gap over 30 minutes. Sessions with fewer than 3 requests or longer than 10 hours were left out. A session belongs to the model that made at least 80% of its cost.
- Prices: every request is priced at Anthropic's list prices from the pricing page (checked November 4, 2026), with the 5-minute cache-write price and Haiku 5.5's long-context price applied per request. These are list-price costs, not what anyone paid.
- Context size: the per-call context figures come from October 31 to November 2.
- Limits: we don't read prompts, so we can't tell what each session was for. A key used in several Claude Code windows at once, or shared by a team, shows up as one bigger session. Requests that failed or were cut off aren't counted.
FAQ
How much does Claude Code cost per hour?
At list prices, the median is about $11 an hour on Opus 5.5, $5 on Sonnet 5.5, $1.40 on Haiku 5.5 and $28 on Fable 5.1. The busiest tenth of sessions cost about 2.5–3 times the median per hour.
How much does one Claude Code session cost?
A median session costs $6.96 on Opus 5.5 and $2.22 on Sonnet 5.5 at list prices. Nine in ten Opus 5.5 sessions stay under about $45.
Why is my Claude Code bill higher than my token estimate?
Each call resends the whole context, often more than 100,000 tokens, and new context is written to the cache at 1.25× the input price. Cache writes are about half of a typical session's cost.
Does Claude Code use prompt caching on an API key?
Yes. In our sample, about 90% of input tokens were cache reads. Without caching, the same sessions would cost about 5–6 times as much.
Is Claude Code cheaper on Sonnet 5.5 than Opus 5.5?
Yes. A median Sonnet 5.5 session costs about a third of a median Opus 5.5 session ($2.22 vs $6.96).