Claude API pricing · read 18 October 2026$0.11 for a call that read 176,040 tokens.
Claude API pricing per million input / output tokens: Opus 5.5 $4 / $20, Opus 5 $5 / $25, Sonnet 5.5 and Sonnet 5 $2 / $10, Haiku 4.5 $1 / $5, Haiku 5.5 $0.10 / $0.50, Fable 5.1 $10 / $50. Cache reads cost a tenth of input or less, and the Batch API halves input and output.
Below: every rate, and what those prices come to on real work — 79,014 Claude Code API calls on one machine, priced line by line.
- median Opus 5 call
- $0.112
- median Opus 5 call
- 44,286 calls
- of the bill was cache reads
- 79%
- of the bill was cache reads
- all Claude models
- per million tokens on Opus 5
- $0.63
- per million tokens on Opus 5
- list input: $5
- the same call without the cache
- 8×
- the same call without the cache
- $0.89
Measured 18 July to 2 October 2026, 56 active days, one developer. Prices from Anthropic's pricing page.
Per million tokens
Claude API pricing per million tokens, every model
Input is fresh text. A cache write stores a prefix for five minutes or an hour; a cache read re-uses it. Output is what the model writes.
| Model | Input | Cache write 5 min | Cache write 1 h | Cache read | Output |
|---|---|---|---|---|---|
| Claude Opus 5.5cache read at 0.05× input | $4 | $5 | $8 | $0.20 | $20 |
| Claude Opus 5same price as Opus 4.5 – 4.8 | $5 | $6.25 | $10 | $0.50 | $25 |
| Claude Sonnet 5.5cache read cut to 0.05× input, 7 Oct 2026 | $2 | $2.50 | $4 | $0.10 | $10 |
| Claude Sonnet 5$2 / $10 is now the standard price | $2 | $2.50 | $4 | $0.20 | $10 |
| Claude Sonnet 4.6and Sonnet 4.5 | $3 | $3.75 | $6 | $0.30 | $15 |
| Claude Haiku 5.5prompts up to 100K tokens; above that, 5× every rate | $0.10 | $0.125 | $0.20 | $0.01 | $0.50 |
| Claude Haiku 4.5 | $1 | $1.25 | $2 | $0.10 | $5 |
| Claude Fable 5.1cache read at 0.025× input; Mythos 5.1 the same | $10 | $12.50 | $20 | $0.25 | $50 |
| Claude Fable 5Mythos 5 the same | $10 | $12.50 | $20 | $1 | $50 |
Cache read is 0.1× input on most models, 0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.1. Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are retired on the Claude API and left out. Mythos models are available only to organizations Anthropic has verified.
How Claude token pricing works
Input
Fresh text the model has not cached: the list input price.Cache write
1.25× input to keep a prefix for five minutes, 2× for an hour.Cache read
0.1× input on most models, 0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.1.Output
What the model writes, thinking included: five times input on every current model.
Claude Opus API pricing
Opus 5.5: $4 / $20 per million input / output tokens, cache read $0.20. Opus 5, and Opus 4.5 to 4.8: $5 / $25, cache read $0.50.
Claude Sonnet API pricing
Sonnet 5.5 and Sonnet 5: $2 / $10; Sonnet 5.5 reads its cache at $0.10, Sonnet 5 at $0.20. Sonnet 4.6 and 4.5: $3 / $15.
Claude Haiku API pricing
Haiku 5.5: $0.10 / $0.50 for prompts up to 100K tokens, five times that above. Haiku 4.5: $1 / $5.
Claude Fable API pricing
Fable 5.1 and Fable 5: $10 / $50; Fable 5.1 reads its cache at $0.25, Fable 5 at $1. Mythos 5.1 and Mythos 5 cost the same, with limited availability.
What changes the price
Batch API: half price
Requests that can wait are billed at 50% of input and output, and it stacks with the cache discount. Not available with fast mode.Long context at the standard price
Claude 4.6 and later run their full 1M-token context at the same per-token rate. The exception is Haiku 5.5: prompts over 100K tokens cost 5× every rate.US-only inference: 1.1×
Pinning inference to the US with inference_geo multiplies every token category by 1.1 on Claude 4.6 and later.Bedrock and Google Cloud regional endpoints: +10%
Regional and multi-region endpoints cost 10% more than global ones, from Sonnet 4.5, Haiku 4.5 and Opus 4.5 on.Fast mode: double
Opus 5.5 at $8 input and $40 output, on the Claude API and Microsoft Foundry on Azure. Opus 5 and Opus 4.8 at $10 and $50, in research preview on the Claude API.The newer tokenizer counts more
Claude 4.7 and later produce about 30% more tokens for the same text than Sonnet 4.6 and earlier, so a cheaper rate on a newer model is not the whole comparison.Web search: $10 per 1,000 searches
On top of the tokens the results add. Web fetch costs only its tokens.
One real call
Claude Opus 5 pricing, on one call
An Opus 5 call from Claude Code, 5 August 2026 — the nearest to the median call in 44,286. The cost has four parts.
- Cache read · the conversation so far, re-sent$0.0880176,040 tokens × $0.50 per million
- Cache write (1 hour) · what this call added to it$0.01971,974 tokens × $10 per million
- Output · what the model wrote$0.0036145 tokens × $25 per million
- Uncached input · everything else$0.00002 tokens × $5 per million
- The call$0.1114
Every turn re-sends the whole conversation. With the cache, those 176,040 tokens cost $0.50 per million instead of $5; without it, the same call would be $0.89.
The same call on every model
The same token counts priced on each model, cheapest first. On cached work the cache-read rate decides, which is why Fable 5.1 at $10 input costs less here than Opus 5 at $5.
- Claude Haiku 5.5$0.0110.10×
- Claude Haiku 4.5$0.0220.20×
- Claude Sonnet 5.5$0.0270.24×
- Claude Sonnet 5$0.0450.40×
- Claude Opus 5.5$0.0540.48×
- Claude Sonnet 4.6$0.0670.60×
- Claude Fable 5.1$0.0910.81×
- Claude Opus 5$0.1111.00×
- Claude Fable 5$0.2232.00×
Relative to Opus 5. Models on the older tokenizer — Sonnet 4.6 and Haiku 4.5 here — would read fewer tokens for the same text, so their rows are upper bounds.
Calculator
Claude API cost calculator
Price a workload the way Anthropic bills it: cache read, cache write, fresh input and output, per call.
Coding agent is a median Opus 5 call from Claude Code on one machine, at the median 890 calls an active day. Median call × median day: a month with long sessions runs higher.
- Per call
- $0.111
- Per day
- $99.14
- Per 30 days
- $2,974
cache read 79% · cache write 18% · output 3% · uncached input 0%
Where the money goes
Claude API cost per model, measured
79,014 calls, $13,922 at list price. Cost per million counts every token a call billed, cached or not.
The bill, by token type
- Cache read78.7%
- Cache write16.3%
- Output5.0%
- Uncached inputunder 0.1%
Uncached input was less than 0.1% of this sample’s list-price cost.
Per model
- Claude Opus 5$0.63 / M
44,286 calls · median call $0.112 · list input $5
- Claude Opus 4.8$0.60 / M
20,993 calls · median call $0.106 · list input $5
- Claude Opus 5.5$0.32 / M
4,653 calls · median call $0.055 · list input $4
- Claude Sonnet 5$0.24 / M
6,830 calls · median call $0.054 · list input $2
- Claude Fable 5$1.35 / M
2,128 calls · median call $0.36 · list input $10
Fable 5.1 reads its cache at a quarter of Fable 5's rate.
In this sample, most of the list-price cost came from reading cached context. Shorter context or a lower cache-read price would reduce that cost. Where the money goes shows tool-call counts from another reading of the same machine.
API or plan
Is the Claude API expensive?
Per call, no: the median Opus 5 call above cost $0.112. For Claude Code all day, yes: 79,014 calls came to $13,922 at list price over 56 active days. Which one fits:
The API fits
an app or a script you ship, work that can wait for the Batch API, and use light enough that a month of tokens at these prices is less than a plan fee.A plan fits
Claude Code most of the day. The calls on this page came to $13,922 at list price over 56 active days; a flat plan billed a fraction of that. The numbers.
Every plan, its limits and the API-key trap that bills a key instead of the plan: Claude Code pricing. Comparing with OpenAI: Codex vs Claude Code.
FAQ
Claude API pricing questions
- How much does the Claude API cost?
- Per million tokens, read 18 October 2026: Opus 5.5 costs $4 input and $20 output, Sonnet 5 $2 and $10, Haiku 4.5 $1 and $5. Cache reads cost a fraction of input: $0.2 on Opus 5.5.
- What does one Claude Code API call cost?
- The median Opus 5 call on this machine cost $0.112 at list price, across 44,286 calls. Most of it was the cache read of the conversation so far, not the new text.
- Is the Claude API cheaper than a Claude plan?
- For light use or an app you ship, often yes. For Claude Code most of the day, no: the 79,014 calls on this page came to $13,922 at list price over 56 active days, far above a flat plan fee.
- How does Claude API prompt caching pricing work?
- A cache write stores a prefix for five minutes or an hour, at 1.25× or 2× the input rate. A cache read re-uses it at 0.1× input on most models, 0.05× on Opus 5.5 and Sonnet 5.5, 0.025× on Fable 5.1.
- Is there a free tier for the Claude API?
- "New users receive a small amount of free credits to test the API," in Anthropic's words; after that every token is billed at the prices above.
- Does the Batch API make Claude cheaper?
- Yes: requests that can wait are billed at 50% of input and output, and the discount stacks with the cache discount. It is not available with fast mode.
Your own calls
See what each of your calls cost, at these prices.
Caprock prices each Claude Code and Codex call from the transcripts already on your disk — per call, per session, per repository and per pull request, at list price. A free desktop app for macOS, Windows and Linux, Apache-2.0, local by default.
Notarized by Apple. One file for Apple silicon and Intel, macOS 13 or later.
Or run it yourself:
Paste in Terminal: it installs Caprock and opens it.Or with Homebrew: brew install --cask dspv/tap/caprock-app
Method: caprock export events, Claude Code assistant turns, one row per API message id, priced at the list prices above with the cache-write lifetime each call reported. Prices read 18 October 2026; check Anthropic's page before you budget.