Claude API pricing per million tokens (2026), every model

Claude API pricing looks simple: a price per million tokens in, a price per million tokens out. What makes bills hard to predict is everything around it: which model, how much context each request carries, whether it is cached, and which tools the model calls along the way.

This guide lists the current prices, explains the modifiers that change them, works through two realistic examples, and ends with the levers that actually move a company's bill. Prices are from Anthropic's pricing page as of October 2026, in US dollars. They change, so check the source before you budget.

Claude API prices per model

Prices are per million tokens (MTok). As a rough guide, a million tokens is about 750,000 English words.

Haiku 5.5's prices apply to prompts up to 100K tokens; longer prompts cost $0.50 input / $2.50 output. Check Anthropic's pricing page for its cache-read rate.

Three things in this table are easy to miss:

• Output costs five times input on every current model. A task that writes long answers costs far more than one that reads a lot and answers briefly.
• Newer is not always pricier. Opus 5.5 is cheaper per token than Opus 5, and Sonnet 5's $2/$10 introductory price became its standard price instead of rising to $3/$15.
• Tokens are not words. Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text. Comparing per-token prices across model generations understates that difference.

Claude 4.6 and later models include the full 1M-token context window at standard prices: a 900k-token request costs the same per token as a 9k-token one. It is still 900k tokens, so long context is a capability, not a discount.

What changes the price

Prompt caching

Most requests repeat a large prefix: the system prompt, tool definitions, a document, the conversation so far. Prompt caching stores that prefix so the next request reads it instead of reprocessing it.

A 5-minute cache pays for itself after a single read. On most models cached tokens also do not count toward your input rate limit, so caching raises throughput as well as cutting cost. For anything with a stable system prompt or repeated documents, it is the single biggest lever.

Batch processing: 50% off

The Message Batches API processes requests asynchronously at half price for both input and output. It suits anything that does not need an answer in seconds: nightly classification, document processing, evaluations, backfills. Batch and caching discounts stack.

Tools

• Web search: $10 per 1,000 searches, plus the tokens of the results the model reads.
• Web fetch: no extra charge, only the tokens of the fetched content. Set max_content_tokens so one large page does not blow up a request.
• Code execution: 1,550 free container-hours per organisation per month, then $0.05 per hour; free when used together with the current web search or web fetch tools.
• Tool definitions add a few hundred input tokens per request for the tool-use system prompt, plus your tool schemas. With dozens of tools or a browser toolset (about 6,600 tokens) this becomes noticeable on every call.

Claude Managed Agents

Anthropic's hosted agent harness bills normal token prices plus $0.08 per session-hour while the session is actually running (idle time waiting for input or approval is free). See Claude Managed Agents for the full picture.

Data residency and fast mode

• US-only inference (inference_geo: "us") costs 1.1x on all token categories for Claude 4.6 and later. Global routing is standard price.
• Fast mode (research preview) for Opus 5.5 costs $8 / $40 per MTok, double the standard rate, for significantly faster output.

Two worked examples

Anthropic's own example: a one-hour agent session on Opus 5 with 50,000 input and 15,000 output tokens costs $0.25 + $0.375 = $0.625 in tokens, plus $0.08 runtime on Managed Agents. With 40,000 of the input tokens read from cache, the token cost falls to about $0.45.

An internal assistant on Sonnet 5.5, our own illustration: 1,000 requests a day, each with 6,000 tokens of system prompt and reference material that is cached, 2,000 tokens of fresh input, and a 500-token answer.

That is roughly $300 a month (cache writes add a little on top). Without caching, the 6M tokens of repeated context would cost $12 a day instead of $1.20, roughly doubling the bill. On Haiku 5.5 the same workload would cost roughly a twentieth; on Opus 5.5 about double.

The pattern holds generally: output and uncached context drive the bill, not the request count.

Spend caps and billing

API usage is billed separately from Claude.ai subscriptions, normally through prepaid credits in the Claude Console. Organisations sit on a usage tier, and each tier has a monthly spend cap:

When you hit the cap, requests fail with a 429 error until the first of the next month or until you move up a tier. You can set your own lower limit on the Billing page, and per-workspace spend and rate limits so one team or one runaway script cannot use up everyone's budget. Rate limits (requests and tokens per minute) also depend on the tier.

Claude API pricing through AWS, Google Cloud and Azure

Claude is also sold through Amazon Bedrock, Google Cloud, Claude Platform on AWS and Microsoft Foundry. Each bills through the cloud provider:

• Bedrock and Google Cloud price Claude themselves. For Claude 4.5 and later models, regional and multi-region endpoints cost 10% more than global ones. This is the route most European companies use to keep inference in an EU region.
• Claude Platform on AWS and Microsoft Foundry are operated by Anthropic and use the same per-token rates as the first-party API, billed in Claude Consumption Units ($0.01 each) through the marketplace.

Batch processing and some features differ by platform, so check the provider's own pricing page before you commit.

How to keep a company's Claude API bill under control

Individual developers optimise prompts. Companies need structure, because the spend is spread across teams, tools and agents nobody has a full view of.

1. Route by task, not by habit. Many requests that go to Opus by default would be just as good on Sonnet or Haiku. A small evaluation set per use case tells you which.
2. Cache by design. Put stable content (instructions, tool definitions, reference documents) first in the prompt and the changing part last, so the prefix can be cached.
3. Batch what can wait. Anything overnight or bulk goes through the Batch API.
4. Cap the context. Limit how much retrieved text, tool output and conversation history goes into each request. Agents that read whole files or web pages are the usual surprise on the bill.
5. Budget per team. Separate workspaces with their own spend limits, so cost is attributable and contained.
6. Watch it weekly. Use the Console's Usage page, or the Usage and Cost API, rather than discovering the trend on the invoice.

Doing that across Claude Code, internal assistants and agents, and across more than one model vendor, is what a gateway is for. Walma AI Hub enforces budgets and model policy per team, forecasts cost per model to month end, and logs every call, running in the customer's own Azure tenant in an EU region. There is a cost calculator on the home page, and the Claude Code pricing guide covers the subscription side of the same question.

Where to go next

• Claude API key: creating keys, key types and team key management.
• Claude Code pricing: subscriptions vs API for developers.
• Claude Managed Agents: what hosted agent sessions cost.