Claude API Pricing: Per-Token Rates, Caching, Batch

The Claude API charges per million tokens, with no monthly fee. Sonnet 5.5 costs $2 in and $10 out, Opus 5.5 costs $4 and $20, Haiku 5.5 costs $0.10 and $0.50, and Fable 5.1 costs $10 and $50. Batch work is half price, and cached input reads cost a tenth or less. Prices checked 2026-10-31 on Anthropic's docs.

Prices checked 2026-10-31

All prices are USD per million tokens (MTok). The source is Anthropic's pricing page in the API docs, cross-checked against claude.com/pricing. Both agree on the current models.

Older models stay on the price list. Sonnet 5 is $2 and $10 (cache read $0.20). Opus 5 and Opus 4.5 through 4.8 are $5 and $25. Sonnet 4.5 and 4.6 are $3 and $15. Haiku 4.5 is $1 and $5. Fable 5 is $10 and $50 with a $1 cache read. Mythos 5.1 matches Fable 5.1 but has limited availability, open only to organizations verified through programs such as Anthropic's Cyber Verification Program.

If you are choosing between current models, the Opus vs Sonnet page covers when the price gap is worth paying. For the model line itself, see Claude models.

Reading the table

Four details trip people up.

1. The newer tokenizer. Anthropic says Claude 4.7 and later models use a tokenizer that produces about 30 percent more tokens for the same text than Sonnet 4.6 and earlier. A cheaper per-token price on a new model does not always mean a cheaper bill on the same prompt. Count tokens on your own text.
2. Haiku 5.5 is priced by prompt length. A request whose prompt is over 100,000 tokens pays the higher row. The prompt length counts all input tokens, including cache reads and writes. Each request is priced on its own.
3. Everyone else gets the full 1M window at one price. Claude 4.6 and later models other than Haiku 5.5 include the full 1M-token context window at standard pricing. A 900K-token request pays the same per-token rate as a 9K one.
4. Sonnet 5 stays at $2 and $10. It launched with introductory pricing through 31 August 2026. The docs now say that price is standard and the planned rise to $3 and $15 will not happen.

Prompt caching

Caching lets the API reuse a prompt prefix across calls. The multipliers are against the base input price.

A 5-minute write pays for itself after one read. A 1-hour write needs two reads. Reads also refresh the cache lifetime.

There are two ways to switch it on. Automatic caching is a single cache_control field at the top of the request, and the system moves the breakpoint as a conversation grows. Explicit breakpoints put cache_control on individual content blocks when you want control over exactly what is cached. Anthropic recommends automatic as the starting point.

Caching also helps your rate limits. For most models only uncached input tokens count toward input tokens per minute, so a good hit rate raises your effective throughput. The docs give an example: a 2,000,000 ITPM limit with an 80 percent hit rate lets you process about 10,000,000 input tokens a minute.

Worked example

Take one request with 50,000 input tokens and 2,000 output tokens on Sonnet 5.5. These are our arithmetic from the table above, not a measurement.

The first caching request costs more than no caching, and every request after it costs about 70 percent less in this shape of job. On Opus 5.5 the same no-cache request is $0.24; on Haiku 5.5 it is $0.006.

Batch processing

The Batch API takes 50 percent off both input and output for asynchronous work. Rates for the current line:

Batch and caching multipliers stack, as do data residency multipliers. Fast mode cannot be used with batch. Use batch for evals, bulk classification, backfills and nightly summaries. Do not use it for anything a person is waiting on.

Modifiers that change the rate

Fast mode applies across the whole context window, including requests over 200K input tokens.

Tool and feature charges

Tool use is mostly just tokens. The tool definitions, tool_use blocks and tool_result blocks all count as input or output. Each model also adds a small hidden system prompt when tools are present: 286 tokens on Opus 5.5, Sonnet 5.5 and Haiku 5.5 with automatic tool choice. The bash tool adds 325 tokens on Claude 4.7 and later, and the text editor tool adds 974.

Server-side tools can add their own charges.

• Web search: $10 per 1,000 searches, plus the tokens of the results. Failed searches are not billed.
• Web fetch: no extra charge beyond tokens. Use max_content_tokens to cap what a page can cost you.
• Code execution: free when used with the current web search or web fetch tools. Otherwise billed by execution time at $0.05 per container hour after a free allowance, with a 5-minute minimum. The docs page gives the allowance as 1,550 free hours per month per organization, and the claude.com pricing page gives it as 50 free hours daily per organization, which is the same allowance counted two ways.
• Managed Agents: standard token rates plus $0.08 per session-hour of active runtime, per claude.com/pricing.

Rate limits and spend caps

Two separate controls limit what you can spend and send.

Spend caps are monthly maxima by tier: Start $500, Build $1,000, Scale $200,000. Custom-tier customers have no cap and agree limits with Anthropic. Hit the cap and requests return HTTP 429 until 00:00 UTC on the first of next month, with no retry-after header, so automatic retries just fail. You can set your own lower cap in the Console, and hitting that returns HTTP 400 instead.

Rate limits are requests, input tokens and output tokens per minute, per model class, enforced as a token bucket that refills continuously.

Fable 5.x has lower limits at every tier: 500,000 input tokens per minute on Start, 1,500,000 on Build and 4,000,000 on Scale. Anthropic places organizations on a tier automatically from usage history. New accounts may start in a lower Evaluation tier while history builds up. A sharp traffic jump can also trigger acceleration limits, so ramp up gradually.

Ways to pay and where else Claude is sold

• Claude Console, direct. Prepaid credits or invoice terms, depending on your agreement.
• Amazon Bedrock and Google Cloud. Claude is sold through both, and the cloud provider invoices you at its own published rates. See their pricing pages for those numbers.
• Claude Platform on AWS and Claude in Microsoft Foundry. Billed through each marketplace in Claude Consumption Units at $0.01 each. Anthropic rates your tokens at the standard per-model rates, applies any discount, then converts.
• Subscription credits. Max ($100), Ultra ($200) and Team ($20 per Standard seat, $100 per Premium seat, pooled and capped at $500 a month) include monthly API credits. You must link one Console organization to claim them, you can link only one, and you cannot change it yourself. Claude, Claude Code and Cowork cannot spend them.

Picking a model by cost

The three-tier shape is simple, and Anthropic's own pages describe each tier.

Price per token is not price per task. Anthropic says newer models need fewer tokens per task than their predecessors, but it does not say how Opus compares with Sonnet on a given job. We did not measure it. Run a slice of your own work on two models and compare the invoice and the retry count.

Our take

The headline is that prices fell. Opus went from $5 and $25 on the 4.x line and Opus 5 to $4 and $20 on 5.5, and Sonnet held at $2 and $10 instead of rising. Cache reads on Opus 5.5 and Sonnet 5.5 dropped to 0.05x. Anyone who modelled costs on Sonnet 4.x numbers is overestimating.

The biggest levers, in order, are model choice, caching and batch. Switching a classification step from Sonnet to Haiku cuts that step's price by 20x on paper. Caching a long system prompt cuts repeated input by 90 to 95 percent. Batch halves what is left. The ones that are easy to miss are the tokenizer change, which can push the same prompt up about 30 percent when you move from a 4.6 model, and the Haiku 5.5 long-prompt threshold, which can make a big-context job cost five times what the headline says.

If your product calls Claude from code, the API is the right meter. If you are one developer using Claude Code, a subscription is likely cheaper than API billing; see Claude Code pricing. For a cross-vendor view, use LLM API prices, and for OpenAI's side, OpenAI API pricing. Subscription tiers are on Claude pricing.

Checklist before you ship

• Count tokens with the real tokenizer on your own prompts.
• Put static content first and add a cache breakpoint; check cache_read_input_tokens in responses.
• Move anything non-interactive to batch.
• Keep Haiku 5.5 prompts under 100K tokens where you can.
• Set a spend limit in the Console below your tier cap, and alert on it.
• Treat 429 with enforced_spend_limit_reached as a stop, not a retry.
• Re-read the pricing page before each budget review. Anthropic changes it.

Common billing surprises

The first cached request costs more. A 5-minute cache write is 1.25x the input price, so the first call in a conversation is dearer than an uncached call. It pays back on the first read. If your traffic is one-off calls with unique prompts, caching adds cost and saves nothing.

Tool definitions count every time. The tools parameter, tool use blocks and tool results are all billed as tokens, and each request with tools carries a built-in system prompt of 286 tokens on the current models. An agent that makes thirty tool calls pays for the tool list thirty times unless it is cached.

Web search is billed per search, not per request. One request can trigger several searches, each at $10 per 1,000.

US-only inference is 1.1x on everything, including cache reads and writes. Add it to the model price before you compare with a vendor that has no residency premium.

Spend caps stop you cold. At the tier cap the API returns a 429 with no retry-after header, so SDK retries only add noise. Set your own limit below the cap and alert before it.

Retired models may linger elsewhere. Anthropic's price list marks Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 as retired except on certain cloud platforms, so check the deprecations page before you pin an old model ID in production.