Claude API Pricing in 2026: what each model costs

Anthropic's current lineup has four models. Prices are in US dollars per million tokens.

Sonnet 5's price needs one clarification. Anthropic launched it at $2 and $10 as an introductory price through 31 August 2026, with an increase to $3 and $15 planned for 1 September. That increase did not happen: $2 and $10 is now the standard price.

Previous-generation models remain available. Claude Opus 4.8, 4.7, 4.6 and 4.5 cost $5 and $25, Claude Sonnet 4.6 and 4.5 cost $3 and $15, and Claude Fable 5 costs $10 and $50. For a new project, Sonnet 5 outperforms Sonnet 4.6 at a lower per-token price.

On recent models the full 1-million-token window is billed at the standard rate: a 900,000-token request costs the same per token as a 9,000-token one.

A token is a fragment of text, often part of a word. Anthropic estimates one token at roughly 4 characters or 0.75 English words.

You pay separately for tokens sent to the model (input: instructions, documents, conversation history) and tokens it generates (output). Output costs five times as much as input on every current model, which is why long answers add up quickly.

Comparing generations has a catch. Claude 4.7 and later models, including Sonnet 5, Opus 5 and Fable 5.1, use a new tokenizer that produces about 30% more tokens for the same text, according to Anthropic. Sonnet 5 at $2 therefore works out to roughly $2.60 measured on the old tokenizer, against $3 for Sonnet 4.6. The saving is real, but smaller than the sticker gap.

When many requests start with the same content (long instructions, a procedures base, a reference contract), prompt caching stops you paying full price for that content every time. The first request writes it to the cache and later ones read it back for a fraction of the price.

The general rule: a write costs 1.25 times the input price for a 5-minute cache and 2 times for a 1-hour cache, and a read costs 0.1 times. Claude Fable 5.1 is the exception, with reads at 0.025 times, or $0.25 per million tokens.

The 5-minute cache pays off after one read, the 1-hour cache after two. Requests sharing the prefix have to arrive within the chosen window: an internal assistant in constant use benefits fully, a job that runs once a day much less.

The Batch API processes requests asynchronously and bills both input and output at 50% of the standard rate. It stacks with prompt caching.

It fits anything that does not need an immediate answer: extracting data from a batch of invoices, sorting requests collected overnight, enriching a contact database. It does not apply to an agent answering a customer live.

Customer support on Haiku 4.5. Anthropic publishes this example: 10,000 support tickets of about 3,700 tokens each cost about $37 with Claude Haiku 4.5.

Invoice extraction on Sonnet 5. Assumption: 1,000 invoices a month, 3,000 input tokens (instructions plus invoice) and 500 output tokens per document. Input costs $6 and output $5, so $11 a month, or $5.50 through the Batch API.

Internal assistant with prompt caching on Sonnet 5. Assumption: 20,000 tokens of procedures placed at the start of every question. Uncached, that prefix costs $0.04 per question, $40 for 1,000 questions. Read from cache, it costs $0.004 per question, $4 for the same 1,000 questions, plus the cost of the writes.

These figures cover API usage only. In a real project, integrating with your tools and maintaining that integration usually costs more, which is what an audit prices before any build. Our page on integrating Claude in the enterprise covers that work.

Anthropic's API routes requests globally by default, at the standard price. Its geography parameter offers a single restriction, US-only processing, billed at 1.1 times the price for Claude 4.6 and later models. There is no European option.

To guarantee processing stays in a specific region, you go through a cloud provider. Amazon Bedrock and Google Cloud offer regional endpoints, priced 10% above their global endpoints. They set their own prices and model availability.

For a GDPR-bound company handling health, legal or candidate data, that 10% premium is rarely the blocker. The question is where the data travels and under which contract. Our guide to the EU AI Act for businesses covers the obligations that apply.

Match the model to how hard the task is:

• Haiku 4.5 for simple, high-volume tasks: sorting requests, classification, extracting well-defined fields.
• Sonnet 5 for most production use: support agents, internal assistants, drafting.
• Opus 5 for complex reasoning and agents that chain many steps.
• Fable 5.1 for the most demanding tasks, when Opus 5 is not enough.

Before building on a model, check its retirement date. Anthropic lists Haiku 4.5 as retiring no sooner than 15 October 2026. An agent built around a model at end of life will need migrating and retesting.

Two separately billed services sometimes come on top: web search costs $10 per 1,000 searches, while web fetch is billed in tokens only. To connect Claude to your software, see MCP in the enterprise, and to compare with OpenAI, Claude vs ChatGPT for business.