Claude API pricing is per token: as of November 2026, Anthropic charges $4 per million input tokens and $20 per million output tokens for Claude Opus 5.5, the model it suggests starting with. Fable 5.1 costs more, Sonnet 5.5 and Haiku 5.5 cost less, and caching, batching and a few options move the bill from there.
Every rate below comes from Anthropic's API pricing documentation, checked on November 4, 2026, in US dollars per million tokens (MTok). Anthropic can change prices at any time, so check the page before you budget.
Current model rates
| Model | Input | Output | Cache hit | Batch input / output |
|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $0.25 | $5 / $25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 | $2 / $10 |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 | $1 / $5 |
| Claude Haiku 5.5, prompt up to 100K tokens | $0.10 | $0.50 | $0.01 | $0.05 / $0.25 |
| Claude Haiku 5.5, prompt over 100K tokens | $0.50 | $2.50 | $0.05 | $0.25 / $1.25 |
Older models keep their own rates. Opus 4.5 through Opus 5 cost $5 in and $25 out, Sonnet 4.6 costs $3 and $15, and Haiku 4.5 costs $1 and $5. Sonnet 5 stays at $2 and $10, because a planned rise to $3 and $15 was dropped.
Four things that change the rate
- Prompt caching. Writing a prefix to the 5-minute cache costs 1.25 times the input rate, a 1-hour write costs twice the input rate, and reading it back costs the cache hit price in the table.
- Batch API. Asynchronous jobs pay half the input and output price, and the discount combines with caching.
- US-only inference. Keeping inference in the US adds 10 percent to every token price on Claude 4.6 and later models.
- Fast mode. Opus 5.5 can answer faster at double the price, $8 in and $40 out. Batches cannot use it.
Two smaller items catch people out. Web search adds $10 per 1,000 searches on top of tokens. And Claude 4.7 and later models use a tokenizer that produces about 30 percent more tokens for the same text, so a lower per-token price on a newer model is not always a lower bill.
A worked request
Take one call to Opus 5.5 with 20,000 input tokens and a 2,000-token reply.
- Input: 20,000 tokens at $4 per million is $0.08.
- Output: 2,000 tokens at $20 per million is $0.04.
- Total: $0.12.
Now say 18,000 of those input tokens are a fixed system prompt that is already cached. They cost $0.0036 at the cache hit rate, the other 2,000 cost $0.008, and the call drops to about $0.05. The first call that writes the cache pays the write rate instead. Send the original uncached call through the Batch API and it costs $0.06.
Claude Code on an API key
Claude Code bills by this same table when you sign in with a Console account or set an ANTHROPIC_API_KEY. Signed in with Pro, Max or Ultra, you pay the plan's flat fee for an allowance instead. Max and Ultra now include monthly API credits, but Anthropic's help center says those credits do not cover interactive Claude Code.
The honest way to choose is to price your real usage. UsageAtlas is a free, open-source desktop tracker for Windows, macOS and Linux that reads local usage from Claude Code and other coding tools. It splits input, output and cache tokens and estimates the cost per model, so you can see what your month would cost at these rates. It needs no account for local tracking and never reads your prompts or responses. Treat its dollar figures as estimates; your invoice is the record.
If you are pricing an app you built on the API, start with the usage page in the Claude Console instead. UsageAtlas tracks coding tools on your computer, not your own application's API calls.