Claude API pricing: what each model costs to run

The Claude API is priced per million tokens, with one rate for the tokens you send and a higher one for the tokens Claude writes back. As of 8 October 2026, Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens, Claude Sonnet 5.5 costs $2 and $10, Claude Opus 5.5 costs $4 and $20, and Claude Fable 5.1 costs $10 and $50. Prompt caching and the Batch API can take a large share off those list prices, and they are where most of the real saving is.

Claude API prices by model

Standard prices per 1 million tokens, from Anthropic's pricing page:

Prices as of 8 October 2026. They change, so check Anthropic's pricing page before you budget, or compare them against OpenAI and Gemini in the API pricing calculator.

Two things stand out. Output costs five times input on every model, so a feature that writes long answers costs far more than its prompt size suggests. And the gap between tiers is wide: Fable 5.1 is a hundred times the price of Haiku 5.5 for the same tokens.

What changes the price you actually pay

The list price is a starting point. Five things move it, two down and three up.

• Prompt caching. If the start of your prompt repeats across requests (a system prompt, instructions, a document everyone asks about), you can cache it. Writing to the cache costs 1.25 times the input price for a five-minute cache, or twice the input price for a one-hour cache. Every later read of it is billed at the cached rate in the table: a tenth of the input price on Haiku 5.5, a twentieth on Sonnet 5.5 and Opus 5.5, and a fortieth on Fable 5.1.
• The Batch API. Requests that can wait are processed asynchronously at half the standard price, input and output both. Batch and caching discounts stack.
• Long prompts on Haiku 5.5. Haiku 5.5's price holds for prompts up to 100,000 tokens. Above that, the whole request is billed at $0.50 input and $2.50 output per million. Sonnet 5.5, Opus 5.5 and Fable 5.1 include their full 1-million-token context at the standard price.
• US-only processing. Pinning inference to the United States applies a 1.1 times multiplier to every token price.
• Server tools. Web search is billed at $10 per 1,000 searches, on top of the tokens the results add to the prompt.

A worked example

Take a support feature that sends 2,000 input tokens per request (a 1,500-token system prompt with instructions and policies, plus 500 tokens of question and retrieved context) and gets back 400 output tokens, across 100,000 requests a month. That is 200 million input tokens and 40 million output tokens.

At list price:

• Claude Haiku 5.5: $20 input plus $20 output, $40 a month
• Claude Sonnet 5.5: $400 input plus $400 output, $800 a month
• Claude Opus 5.5: $800 input plus $800 output, $1,600 a month
• Claude Fable 5.1: $2,000 input plus $2,000 output, $4,000 a month

Now cache the 1,500-token system prompt on Sonnet 5.5. With steady traffic it is read from the cache on almost every request:

• Cached input: 150 million tokens at $0.10 per million, $15
• Fresh input: 50 million tokens at $2 per million, $100
• Output: 40 million tokens at $10 per million, $400
• Total: about $515 a month, plus a few dollars of cache writes

That is roughly a third off, from one change. Notice what is left: output is now almost 80% of the bill. Once caching is in place, the next saving comes from shorter answers, not a shorter prompt.

Choosing a Claude model on cost

Price per token is only half the comparison. A cheaper model that needs two attempts, or gets one answer in ten wrong, can cost more than a dearer one that gets it right first time. The way I choose is to build a test set from real questions, run it on the cheapest model that might work, and move up a tier only where the results fall short.

In practice, that usually means a mix rather than one model: a small model for routing, classification and extraction, and a larger one only for the steps that need it. I put the model behind an interface in every product I build, so changing it later is a configuration change rather than a rewrite.

Claude vs OpenAI and Gemini pricing

At matching tiers the list prices are close. Claude Sonnet 5.5 and OpenAI's GPT-6.1 Sol both cost $2 input and $10 output per million tokens, and Claude Haiku 5.5 and GPT-6 Luna both cost $0.10 and $0.50. Google's Gemini 3.1 Pro (preview) is $2 input and $12 output, slightly dearer on output.

The real differences are in the details above: how cheap a cache hit is, where long-prompt prices step up, and how many tokens each model needs to do your task well. Those are worth measuring on your own requests. The Gemini API pricing guide covers Google's side, and what an AI feature costs to run has the full cross-vendor table.

Measure before you switch

Every lever here is a guess until you know what one request costs today and which part of it is expensive. On Evoriqa, my multi-tenant support platform, every AI action draws from one credit balance, so each workspace sees what its AI use costs. On an AI audit reviewer I built for a CPA firm, logging tokens on every model call put a review at a measured $0.24-0.28, and found a setting that produced more findings at 63% of the cost.

If your Claude bill is growing faster than your usage, an LLM cost and RAG quality audit measures cost per request on your real traffic and ranks the changes by what they save.