Short answer: Anthropic’s published Claude API prices are charged per million tokens, with separate rates for input and output. Claude Opus 5.5 is $4 per million input tokens and $20 per million output tokens; Sonnet 5.5 is $2/$10; and Haiku 5.5 is $0.10/$0.50 for prompts up to 100,000 tokens. These are API rates—not a per-token price for a Claude chat subscription. Haiku pricing above 100,000 prompt tokens is not specified here, so confirm the applicable rate before budgeting that workload. Prices below are in USD and reflect Anthropic’s published documentation available on November 10, 2026. (Anthropic pricing)
Claude API token prices
| Claude model | Input price per 1 million tokens | Output price per 1 million tokens | Pricing note |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | Anthropic’s model page lists these API rates. (Model page) |
| Claude Opus 5.5 | $4 | $20 | Standard model pricing; separate rates apply to caching and batch processing. (Anthropic pricing) |
| Claude Sonnet 5.5 | $2 | $10 | Standard model pricing; separate rates apply to caching and batch processing. (Anthropic pricing) |
| Claude Haiku 5.5 | $0.10 | $0.50 | Rates shown for prompts up to 100,000 tokens. Above that prompt size, the applicable rates and threshold treatment are not specified here; confirm them before estimating. (Anthropic pricing) |
Anthropic’s pricing documentation states that prices are in USD. Its published list-price card is effective October 7, 2026; model-specific documentation also lists Fable 5.2 at $8/$40 per million tokens. Availability and rates can differ by platform or inference scope, so confirm the exact model and billing channel you will use before forecasting spend. (List-price card)
Estimate a production API bill
For a basic estimate, calculate input and output charges separately:
Estimated token cost = (input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)
Then multiply by the number of requests in the billing period. Use measured or realistically sampled token counts; an estimate based only on characters can be off because token counts depend on the text and model.
Example using Sonnet 5.5: Assume 1,000 API calls per day, each averaging 2,000 input tokens and 1,000 output tokens, for 30 days. Using the published rates of $2 per million input tokens and $10 per million output tokens: (Anthropic pricing)
- Input: 1,000 × 2,000 × 30 = 60 million tokens; 60 × $2 = $120.
- Output: 1,000 × 1,000 × 30 = 30 million tokens; 30 × $10 = $300.
- Estimated total: $420 for 30 days, or $0.014 per call at that assumed token mix.
This is a base-token estimate, not a guaranteed invoice. It excludes prompt-cache writes and reads, batch pricing adjustments, separately priced tools or features, and any platform-specific billing differences. Anthropic’s pricing documentation lists cache and feature pricing separately. (Anthropic pricing)
API pricing versus Claude subscriptions
Claude chat plans and Claude API usage are different billing channels: a paid chat subscription is not itself a per-token API rate. Anthropic says API usage is billed separately, while noting that Max, Ultra, and Team plans include monthly credits for the Claude Platform. Those credits are an allowance toward API use, not a conversion of the subscription price into tokens; check the plan’s eligibility and credit terms before counting them in a budget. (Anthropic Support)
Production factors that can change the estimate
- Input/output mix: Long prompts and long completions can have very different costs because input and output rates differ. Estimate both sides from representative production requests.
- Prompt caching: Reused prompt prefixes may be billed at separate cache-write and cache-read rates. Anthropic’s documented cache-read multiplier is 0.1× the standard input rate: that works out to $0.40 per million cache-read tokens for Opus 5.5 and $0.20 for Sonnet 5.5 at their standard input rates. For eligible batch requests, the 50% batch discount can stack with the cache-read modifier, yielding $0.20 and $0.10, respectively, for those cache-read tokens. Cache writes have separate rates; whether caching reduces your total depends on the workload and feature requirements. (Prompt caching; Anthropic pricing)
- Asynchronous batch work: Anthropic says its Message Batches API can reduce costs by 50% for eligible work that does not need an immediate response. Batch and prompt-caching modifiers can stack; apply them only to qualifying usage rather than as a blanket discount on every request. (Batch processing)
- Model choice: A lower per-token rate does not by itself prove a model is a suitable substitute. Test quality, latency, and token usage on the actual workload, then compare total cost for the task.
For a safer production forecast, measure a representative sample of calls, separate input, output, and cached tokens, apply the rate for the exact model and API channel, and scale by expected request volume.
Sources
Pricing - Claude Platform Docs
Claude Fable 5.2 - Claude Platform Docs
Anthropic List Prices — 2026-10-07
Why is Claude API usage billed separately from my paid Claude plan?