Published 10 June 2026. Last updated: 12 September 2026. Prices below are from Anthropic's pricing documentation, checked that day.
Claude API pricing is what Anthropic charges per token when you call Claude models from your own code, as opposed to a flat Claude subscription. You pay separately for input, output, cache writes and cache reads, and for agent-style workloads the last one decides most of the bill.
For subscriptions and their limits, see Claude plans: Pro vs Max. For free ways onto the API, see Claude free credits.
Claude API prices per model (official, September 2026)
Current models, per million tokens:
| Model | Input | Output | Cache read | 5-min cache write | 1-hour cache write |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | $1 | $5 | $0.10 | $1.25 | $2 |
| Claude Sonnet 5 | $2 | $10 | $0.20 | $2.50 | $4 |
| Claude Opus 5 | $5 | $25 | $0.50 | $6.25 | $10 |
| Claude Fable 5.1 | $10 | $50 | $0.25 | $12.50 | $20 |
Claude Mythos 5.1 is priced the same as Fable 5.1 but has limited availability. Source: Anthropic's pricing documentation, checked 12 September 2026.
Older models that are still available:
| Model | Input | Output | Cache read |
|---|---|---|---|
| Claude Fable 5 | $10 | $50 | $1 |
| Claude Opus 4.8, 4.7, 4.6, 4.5 | $5 | $25 | $0.50 |
| Claude Sonnet 4.6, 4.5 | $3 | $15 | $0.30 |
Opus 4.1, Sonnet 4 and Haiku 3.5 are retired on the Claude API and remain only on Amazon Bedrock and Google Cloud; Opus 4 remains only on Google Cloud.
What changed since June?
- Sonnet 5 stays at $2/$10. It launched at that price as introductory pricing through 31 August 2026, with a rise to $3/$15 planned for 1 September. Anthropic cancelled the rise, so the newest Sonnet is now cheaper than Sonnet 4.6.
- Fable 5.1 cut the cache read price by three quarters. A cache hit on Fable 5.1 costs 0.025x the input price ($0.25 per million tokens) against 0.1x ($1) on Fable 5, while input and output prices stayed the same.
- Opus 5 costs the same as Opus 4.8, at $5/$25.
- The tokenizer figure is about 30%. Claude 4.7 and later models produce about 30% more tokens for the same text than Sonnet 4.6 and earlier, by Anthropic's own estimate. The June version of this page said up to 35%.
What does an agent workload actually cost?
For an agent that works inside a large context, such as Claude Code, the headline input price barely matters. My own Claude Code statistics over several months of daily work split like this: about 98.6% of tokens were cache reads, 1.2% cache writes, 0.25% output and a rounding error of fresh input. Applying that mix to each model's prices gives the cost of one million tokens of that kind of work:
| Model | Cost per 1M tokens of this mix (5-min cache writes) | With 1-hour cache writes |
|---|---|---|
| Claude Haiku 4.5 | $0.13 | $0.14 |
| Claude Sonnet 5 | $0.25 | $0.27 |
| Claude Fable 5.1 | $0.52 | $0.61 |
| Claude Opus 5 | $0.63 | $0.68 |
| Claude Fable 5 | $1.26 | $1.35 |
Two things follow. The cache read price is most of the bill: on every model except Fable 5.1 it is about 78% of the total for this mix. And because Fable 5.1's cache reads are so cheap, Fable 5.1 comes out cheaper than Opus 5 on this workload even though its list price is twice as high. For a chat app with short contexts the order flips back, so work out your own mix before choosing a model on price.
The token proportions are measured from my own usage; the prices are Anthropic's; the result is arithmetic, not a billing statement.
How do prompt caching and the Batch API cut the bill?
Prompt caching bills repeated context at a fraction of the input price. The multipliers, from Anthropic's documentation:
| Cache operation | Price | How long the cache lives |
|---|---|---|
| 5-minute cache write | 1.25x base input | 5 minutes |
| 1-hour cache write | 2x base input | 1 hour |
| Cache read | 0.1x base input (0.025x on Fable 5.1 and Mythos 5.1) | Same as the write |
A 5-minute write pays for itself after one cache read; a 1-hour write after two.
The Batch API processes requests asynchronously at 50% off input and output: Opus 5 drops to $2.50/$12.50, Sonnet 5 to $1/$5, Haiku 4.5 to $0.50/$2.50 and Fable 5.1 to $5/$25. Caching multipliers and data residency multipliers stack on top of the batch price.
What other charges are there?
| Item | Price |
|---|---|
| Fast mode (Opus 5 and Opus 4.8, research preview) | $10 input / $50 output per million tokens, Claude API only, not with Batch |
US-only inference (inference_geo: "us", Claude 4.6 and later) | 1.1x on every token category |
| Long context (Claude 4.6 and later) | Full 1M-token window at standard per-token prices |
| Web search | $10 per 1,000 searches, plus result tokens |
| Web fetch | No charge beyond tokens |
| Code execution | Free with web search or web fetch; otherwise 1,550 free hours a month per organization, then $0.05 per hour per container |
Declaring any tool also adds a hidden system prompt: 286 input tokens on Opus 5, 354 on Sonnet 5 and 496 on Haiku 4.5 with tool_choice set to auto.
How do I estimate a Claude API bill?
- Estimate input and output tokens per request, and add about 30% if you measured the text with an older tokenizer.
- Multiply by the model's input and output prices.
- If a large prefix repeats (system prompt, documents, conversation history), price that part at the cache read rate after the first call.
- If the work does not need an immediate answer, halve it with the Batch API.
Example: a Sonnet 5 chat turn with 2,000 input tokens and 500 output tokens costs $0.004 + $0.005 = $0.009 before caching.
Claude API or a Claude subscription?
A Claude subscription is a flat monthly price with usage limits; the API bills every token with no ceiling and no limit to wait out. If you use Claude yourself, in the apps or Claude Code, start with the plans and their limits. If you are building a product on Claude, the API is the only option, and the tables above are your cost model. For Claude Code specifically, see Claude Code pricing.
Frequently asked questions
How much does the Claude API cost in 2026?
Per million input and output tokens in September 2026: Claude Haiku 4.5 costs $1 and $5, Sonnet 5 $2 and $10, Opus 5 $5 and $25, and Fable 5.1 $10 and $50. Cache reads cost a tenth of the input price (a fortieth on Fable 5.1), and the Batch API takes 50% off.
Which Claude model is cheapest?
Claude Haiku 4.5, at $1 per million input tokens and $5 per million output tokens. Sonnet 5 is the cheapest current mid-tier model at $2 and $10, which is also cheaper than the older Sonnet 4.6 at $3 and $15.
Did Claude Sonnet 5 go up to $3/$15 in September 2026?
No. Sonnet 5 launched at $2/$10 as introductory pricing through 31 August 2026, with an increase to $3/$15 scheduled for 1 September. Anthropic cancelled the increase, and $2/$10 is now the standard price.
How does prompt caching reduce Claude API cost?
A cache read costs 0.1x the base input price on most models and 0.025x on Fable 5.1. Writing to the cache costs 1.25x for a 5-minute cache or 2x for a 1-hour cache. A 5-minute write pays for itself after one read, a 1-hour write after two.
Is there a free Claude API tier?
No. Anthropic has no free tier on the Claude API. The free routes are credit programs, such as Anthropic's startup program. For zero-cost inference on other models, Google Gemini and Groq still have free API tiers in September 2026.
Why does Claude use more tokens than older models for the same text?
Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text, according to Anthropic. The exact increase depends on the content, so compare providers on your own real prompts, not on the per-token price alone.
How much does Claude web search cost on the API?
Web search costs $10 per 1,000 searches plus the tokens of the results it adds to the conversation. Web fetch has no extra charge beyond tokens, and code execution is free when used together with web search or web fetch.
