LEARN / CLAUDE API COST

How much does the Claude API cost? 0.04¢ to 4.0¢ per request, model by model

PUBLISHED 2026-11-10 · BY Llorenç Ballester — Chief Burn Officer

Anthropic publishes Claude's prices per million tokens. Nobody sends a million tokens; they send requests. This page turns the price list into what a request actually costs, on every Claude model, and shows the two things that move it most: cached context and output length. Every figure is computed from our live price board, so it moves when prices move.

What a request costs, model by model

Three requests you are likely to send, priced at list on every Claude model: a chat reply (1,500 in / 500 out), a long document summarised (30,000 in / 1,000 out), and one turn of a coding agent (50K of context, all but 2K read from cache, 1,000 out).

Per million tokens, input/output, in brackets. List prices, no batch discount. The agent turn uses each model's published cached-read rate. Prices last confirmed 2026-11-10. Every model links to its own price page.

Two things stand out. Newer is not dearer: Claude Haiku 5.5 costs less than the Haiku it replaces, and Claude Opus 5.5 less than Opus 5. And the agent column dwarfs the chat column on every model, which is the next section.

Claude's cost per token

Per-million prices hide how small a token is, and how different the cache lines are:

From our Claude price page, recomputed on every deploy. “—” means no cached rate is published, not that caching is free.

The cache decides an agent's bill

Models are stateless, so every turn of a conversation or an agent re-sends the whole context. Claude lets you cache that context and read it back for a fraction of input: the last column above, from 0.1x down to a fortieth on the newest models. On Claude Sonnet 5.5, the agent turn above costs $0.0188 with the context cached and $0.11 without, 5.9x more.

Writing to the cache is the part people miss. Anthropic bills a write at 1.25x input for a five-minute lifetime and 2x for an hour. If the context is not read back before it expires, you paid a premium for nothing. Below a minimum prompt length nothing caches at all, silently: the cache minimum explains it, and five minutes or one hour shows which lifetime pays on real agent traffic.

What Claude Code adds

Claude Code is the agent case at scale. In five months of one developer's logs, priced at list, the uncached prompt was 1% of the bill; cache reads and one-hour cache writes were most of the rest. The full breakdown is in where a Claude Code bill actually goes. To see your own split, run npx tokenburnindex audit: it reads the logs Claude Code already keeps on your disk and nothing leaves your machine.

Three more things that move the price

• — Output length. Every Claude model bills output at 5x input, so a long answer costs more than a long question. Why output costs more.
• — Batch. Requests that can wait go through the Batch API at half price: Claude Sonnet 5.5 drops to $1 in / $5 out.
• — Long prompts on Claude Haiku 5.5. Past 100K prompt tokens the whole request reprices from $0.1/$0.5 to $0.5/$2.5.
• — The tokenizer. Anthropic's newer models turn the same text into more tokens, so the same prompt can cost more with no price change. The tokenizer tax.

Price your own prompt

Averages are for articles. Paste a real prompt into the Burnmeter for its token count and its cost across 43 models, or compare any Claude model head to head with GPT, Gemini and the rest. If you already use the API, drop your usage export on the audit to see where last month went. It runs in your browser.

Frequently asked questions