Claude API Pricing (2026): Per-Token Costs and Examples

The Claude API charges per token, and the bill depends on three choices: which model you call, how much of each prompt you can cache, and whether the work can wait for a batch. This guide lists the current per-token prices, explains every discount and surcharge, and walks through four worked cost examples you can adapt.

All prices were checked on October 24, 2026 against Anthropic’s official pricing page and claude.com/pricing. Prices are in USD and change; confirm before you budget. If you’re paying for the Claude apps or Claude Code rather than the API, see our Claude pricing overview and Claude Code pricing guide instead.

Current model prices (per million tokens)

MTok means one million tokens. A token is roughly a short word or part of a word.

Model	Input	5-min cache write	1-hour cache write	Cache read	Output
Claude Fable 5.1	$10	$12.50	$20	$0.25	$50
Claude Opus 5.5	$4	$5	$8	$0.20	$20
Claude Sonnet 5.5	$2	$2.50	$4	$0.10	$10
Claude Haiku 5.5 (prompts up to 100k tokens)	$0.10	$0.125	$0.20	$0.01	$0.50
Claude Haiku 5.5 (prompts over 100k tokens)	$0.50	$0.625	$1	$0.05	$2.50

Older models are still listed, often at higher prices. A few examples from the same page: Opus 5 and Opus 4.5 through 4.8 are $5 input and $25 output; Sonnet 4.5 and 4.6 are $3 and $15; Haiku 4.5 is $1 and $5. If you are still on one of those, moving to the 5.5 generation is usually a price cut as well as an upgrade. Claude Mythos 5.1 is priced like Fable 5.1 but is available only to organisations verified through Anthropic’s access programs.

Which model should you price first?

The pricing page’s cost advice is: Haiku for simple tasks, Sonnet for most production workloads, Opus for the most complex reasoning. The models overview suggests starting quality tests on Opus 5.5 and reserving Fable 5.1 (2.5x Opus’s price) for demanding, long-horizon work. A practical approach: prove quality on Opus 5.5, then test whether Sonnet 5.5 or Haiku 5.5 passes the same checks for the high-volume parts. All four current Claude models have a 1M-token context window and 128K max output.

Two details that change your token count

• Tokenizer. Claude 4.7 and later models use a newer tokenizer that, per the pricing page, produces about 30% more tokens for the same text. Measure with the token counting endpoint rather than reusing estimates from older models.
• Haiku 5.5’s 100k threshold. A request whose prompt exceeds 100,000 tokens (cache reads and writes included) pays the higher Haiku rates for that whole request. Other 4.6-and-later models include the full 1M-token context window at standard prices.

Prompt caching: the biggest lever

Prompt caching stores a prompt prefix (system prompt, documents, tool definitions, conversation history) so later requests read it cheaply instead of reprocessing it.

Cache operation	Price multiplier	Lifetime
5-minute cache write	1.25x base input	5 minutes, refreshed on each hit
1-hour cache write	2x base input	1 hour
Cache read	0.1x base input (0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1)	Same as the write

The break-even is fast: a 5-minute cache pays for itself after one read, a 1-hour cache after two. You can enable it with a single top-level cache_control field (automatic caching) or place breakpoints on specific content blocks.

{
  "model": "claude-sonnet-5-5",
  "max_tokens": 1024,
  "cache_control": { "type": "ephemeral" },
  "system": "You are the support assistant for Acme... (long policy text)",
  "messages": [{ "role": "user", "content": "How do I reset my password?" }]
}

Check the model ID and the minimum cacheable prompt length for your model in the prompt caching docs before relying on the example above. Then confirm caching is working by reading cache_read_input_tokens in the response’s usage block.

Batch API: 50% off for work that can wait

The Message Batches API processes requests asynchronously at half the price of both input and output tokens.

Model	Batch input	Batch output
Claude Fable 5.1	$5	$25
Claude Opus 5.5	$2	$10
Claude Sonnet 5.5	$1	$5
Claude Haiku 5.5 (up to 100k)	$0.05	$0.25

Good fits: nightly classification, document summarisation, evals, data extraction and content backfills. Caching multipliers stack on top of the batch discount.

Surcharges and extra fees

Item	Price (as of Oct 24, 2026)
US-only inference (inference_geo: "us"), Claude 4.6+	1.1x on all token categories
Fast mode, Opus 5.5	$8 input / $40 output per MTok; not available with Batch
Web search tool	$10 per 1,000 searches, plus tokens for the results
Web fetch tool	No extra fee; you pay for fetched content as tokens
Code execution	1,550 free container-hours per organisation per month, then $0.05 per hour per container; free when used with the current web search or web fetch tools
Claude Managed Agents	Tokens at model rates plus $0.08 per session-hour of running time
Tool use overhead	A hidden tool-use system prompt, for example 286 tokens on Opus 5.5 and Sonnet 5.5, plus your tool schemas

Cloud platforms (Amazon Bedrock, Google Cloud) publish their own prices, and regional endpoints there carry a 10% premium over global ones.

Worked cost examples

All examples use list prices from the tables above and ignore tax. Your token counts will differ; the point is the method. To run the same maths on your own numbers, use our Claude API cost calculator.

1. Support chatbot on Sonnet 5.5

10,000 conversations a month. Each request sends a 2,000-token system prompt plus 1,000 tokens of new conversation, and returns 500 tokens.

	Without caching	With caching
System prompt (20M tokens)	20 × $2 = $40	20 × $0.10 = $2
New input (10M tokens)	10 × $2 = $20	10 × $2 = $20
Output (5M tokens)	5 × $10 = $50	5 × $10 = $50
Total	$110	about $72

Cache writes add a few cents whenever the 5-minute cache expires. Notice that output dominates once input is cached: trimming verbose answers is the next saving.

2. Bulk summarisation on Haiku 5.5 with Batch

100,000 documents, each 4,000 tokens in and 300 tokens out (well under the 100k threshold).

• Standard: 400M input × $0.10 = $40, plus 30M output × $0.50 = $15. Total $55.
• Batch: 400M × $0.05 = $20, plus 30M × $0.25 = $7.50. Total $27.50.

3. One coding-agent task on Opus 5.5

An agent makes 40 requests. Each carries about 60,000 tokens of context and returns 2,000 tokens. Agent loops resend the conversation every turn, so caching does most of the work. Assume 90% of input is read from cache and 10% is newly written to a 5-minute cache.

• Cache reads: 2.16M × $0.20 = $0.43
• Cache writes: 0.24M × $5 = $1.20
• Output: 0.08M × $20 = $1.60
• Total about $3.23, versus about $11.20 with no caching (2.4M × $4 + $1.60).

This is why tools like Claude Code cache automatically. Our Claude Code usage and cost guide covers subscription versus API billing for agent work.

4. Ten questions about a 500k-token document on Fable 5.1

• No caching: each question costs 0.5M × $10 = $5 input, plus about 5,000 output tokens × $50/MTok = $0.25. Ten questions: $52.50.
• 5-minute caching: the first call writes the document (0.5M × $12.50 = $6.25) plus $0.25 output. Each of the nine follow-ups reads it (0.5M × $0.25 = $0.125) plus $0.25 output. Total about $9.88, if each question arrives within five minutes of the last.

How to keep the bill predictable

1. Set a workspace spend limit in the Claude Console for each project, so one runaway loop can’t drain the organisation’s credits.
2. Log usage from every response. Input, output, cache read and cache write tokens are all reported per request.
3. Cap max_tokens to what the task needs. Output is the most expensive token type on every model.
4. Cache the stable prefix. Put the system prompt, tool definitions and reference documents first, and the changing user turn last.
5. Move offline work to Batch. Anything that doesn’t need an answer within seconds is a candidate.
6. Route by difficulty. Send easy requests to Haiku 5.5 and escalate only when a check fails.

Well-structured prompts also cost less, because they get the right answer in fewer turns; our Claude prompting best practices shows how. If you build agents on the API, the MCP servers guide explains how connected tools add to your token count.

Subscription credits and enterprise discounts

• Max and Team plans include API credits. According to Anthropic’s API credits page, Max 5x includes $100 a month, Max 20x $200, and Team up to $500 pooled ($20 per Standard seat, $100 per Premium seat). They cover the Claude API, Managed Agents, the Agent SDK and the Console playground, but not Claude Code, and they expire each billing cycle. Free, Pro and Enterprise plans aren’t eligible.
• Volume discounts are negotiated case by case with Anthropic sales, and academic or research discounts may be available.

Where API costs fit in your Claude kit

If you’re assembling a toolkit around Claude, API spend is usually the only recurring cost: skills, subagents and hooks are free features, but they all run on tokens. Our Claude kit hub and What is a Claude kit? explain how the pieces fit, and ccusage helps you see where tokens go when you run Claude Code locally.