How Claude API pricing works
Anthropic bills the Claude API per token, with separate prices for the text you send (input) and the text Claude writes (output). Output costs five times as much as input on every current model. There are four current Claude models, and Anthropic positions each for a different kind of work:
- Claude Fable 5.2 ($8.00 / $40.00): the most capable model, for demanding reasoning and long-horizon agentic work. Released 9 November 2026, 20% cheaper than Fable 5.1; what changed.
- Claude Opus 5.5 ($4.00 / $20.00): long-running agentic coding and knowledge work.
- Claude Sonnet 5.5 ($2.00 / $10.00): what Anthropic calls the best combination of speed and intelligence.
- Claude Haiku 5.5 ($0.10 / $0.50): high-volume, latency-sensitive work such as classification, extraction and routing.
All four have a 1M-token context window. Anthropic charges the same per-token rate across the whole window on Fable, Opus and Sonnet, so a 900,000-token request costs the same per token as a 9,000-token one. Haiku 5.5 is the exception, covered below.
Haiku 5.5's 100,000-token threshold
If a Claude Haiku 5.5 request's prompt is over 100,000 tokens, the whole request is billed at $0.50 / $2.50 per 1M, five times the normal rate, not just the tokens over the line. Cached tokens count towards the 100,000. So a prompt of 99,000 tokens with a 1,000-token answer costs about $0.010, while 101,000 tokens with the same answer costs about $0.053. For long documents, split the work into chunks under 100,000 tokens, or compare the cost with Claude Sonnet 5.5, which charges the same rate across its whole 1M-token window.
Prompt caching and batch
Caching is where most of the savings are for agents and chat apps, because they resend the same system prompt, tools and conversation on every call. Anthropic bills it in two parts:
- Writing to the cache costs more than normal input: 1.25× the input price for a cache that lasts 5 minutes, or 2× for one that lasts an hour.
- Reading from the cache costs a fraction of input: 0.1× on Haiku 5.5 and older models, 0.05× on Opus 5.5 and Sonnet 5.5, and 0.025× on Fable 5.2 and 5.1. On Sonnet 5.5 that is $0.10 per 1M tokens instead of $2.
When a cache write pays for itself: by Anthropic's own figures, a 5-minute write pays off after one cache read and a 1-hour write after two. Use the 1-hour cache when calls arrive minutes apart rather than seconds apart, as with a person working through an agent session with pauses. These multipliers stack with the Batch API discount and with US-only processing, so a batched request that reads from cache gets both discounts.
For a coding agent doing 400 tasks a month (about 750M input tokens, 8M output), Sonnet 5.5 costs $1,580 a month with no caching and $369 with 85% of input served from cache, which is typical for agents. Opus 5.5 comes to $738 for the same work. These figures leave out cache-write charges, which add a little on top.
The Batch API takes 50% off input and output for requests you can wait hours for, such as evaluations, data processing or overnight reports.
Is there a free tier?
Only a little. Anthropic's pricing page says new API users receive a small amount of free credit to test the API, without saying how much; after that you pay per token, in advance. Enterprise teams can ask Anthropic's sales team for an extended trial. The Claude apps at claude.ai have a free plan, but it doesn't include API access. Free tiers from every provider.
Subscription or API?
A Claude subscription and the Claude API are billed separately. Pro ($20 a month), Max ($100) and Ultra ($200, the plan formerly called Max 20x) cover the Claude apps and Claude Code, within usage limits. They don't cover API calls from your own code, with one exception: Max includes $100 a month in Claude Platform API credits and Ultra $200, and Team plans get up to $500 a month, pooled across seats.
For one person using Claude every day, a subscription is usually cheaper than paying per token. For an app, an automation or anything serving other people, you need the API. Our subscription vs API calculator works out the break-even for your usage, and the Claude Code pricing page covers the coding plans in detail.
Claude on AWS, Google Cloud and Azure
You can also buy Claude through your cloud account, and what you pay depends on who runs the service:
| Where you buy it | Who sets the price | What it costs |
|---|---|---|
| Claude Platform on AWS, Claude on Google Cloud (Anthropic managed) | Anthropic | Anthropic's standard per-model rates, billed through the cloud marketplace |
| Claude in Microsoft Foundry | Anthropic | Standard rates converted to Claude Consumption Units at $0.01 each, on your Azure bill |
| Amazon Bedrock, Claude on Google Cloud (Google managed) | The cloud provider | Its own published prices. Regional and multi-region endpoints cost 10% more than global ones |
Keeping data in the US costs 1.1 times the standard rate on the Claude API, on the Anthropic-managed cloud versions and on Foundry's US Data Zone deployments. Through the Anthropic-run options, then, the price per token matches buying direct; buying that way mainly helps when you want one bill or already have committed spend there. On Bedrock and Google-managed Vertex, check the cloud's own price list, because regional pricing is set independently.
Which Claude model should you use?
- Most coding, agents and everyday work: start with Sonnet 5.5. It costs half as much as Opus 5.5 on every rate.
- Complex, long-running tasks that need careful judgment: Opus 5.5. Test it against Sonnet on your own tasks; the extra cost pays off only when it gets things right more often.
- High-volume, simpler jobs (classification, extraction, summaries, routing, subagents): Haiku 5.5, at a twentieth of Sonnet's price. Watch for prompts over 100,000 tokens.
- The hardest problems, where cost matters less: Fable 5.2, at twice the price of Opus 5.5. Anthropic itself suggests starting with Opus 5.5 and moving up only when a task needs it: Fable 5.2 vs Opus 5.5.
A common pattern is to plan with Opus or Sonnet and hand routine steps to Haiku; the agent cost calculator shows what caching saves a coding agent on each. Against other providers: Sonnet 5.5 vs GPT-6.1 Sol (the same per-token price), Opus 5.5 vs GPT-6 Astra and Haiku 5.5 vs GPT-6 Luna.
Availability: Anthropic suspended access to Claude Fable 5 and Mythos 5 from 12 June to 1 July 2026 after a US export-control directive. What happened, and what it means for planning. If an app depends on the top model, keep a fallback to Opus or Sonnet configured.
Other charges
- Web search costs $10 per 1,000 searches, plus normal token costs for the search results Claude reads.
- Code execution is free when used with web search or web fetch. Otherwise each organisation gets 1,550 free container hours a month, then $0.05 per hour.
- US-only processing: asking for inference to run only in the United States adds 10% to every token price on current models. The default global routing uses standard prices.
- Fast mode generates output faster at twice the standard rate: $16 / $80 per 1M tokens on Fable 5.2 and $8 / $40 on Opus 5.5.
Claude is also sold through Amazon Bedrock and Google Cloud Vertex AI, which set their own prices; this page covers Anthropic's own API.
Sources: Anthropic API pricing, Claude Team plan billing, Anthropic, 7 October 2026. API prices, caching, fast mode and processing rules checked 10 November 2026; plan credits checked 13 October 2026.
