November 21, 2026
Anthropic API Pricing in Plain English: Every Claude Model, Per Million Tokens
Anthropic API Pricing in Plain English: Every Claude Model, Per Million Tokens

Title tag: Anthropic API Pricing: Every Claude Model, Per Million Tokens
Slug: anthropic-api-pricing
Anthropic’s API is billed per token. Its current lineup ranges from Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens to Fable 5.2 at $8 in and $40 out. Everything else is a variation on that.
Key takeaways
- Current standard input rates range from $0.10 to $8 per million tokens; output rates range from $0.50 to $40.
- A Sonnet 5.5 cache read costs $0.10 per million tokens, compared with $2 for standard input. Caching helps when you reuse a large prompt prefix.
- Batch processing cuts input and output token rates by 50% for work that can run asynchronously.
- The biggest cost decision is usually model choice. Match the model to the task, then consider caching and batch processing.
Anthropic API pricing per million tokens
Prices below are USD per million tokens (MTok) for standard processing. Haiku 5.5 has different rates depending on whether a request’s prompt is above 100,000 tokens.
| Model | Input / MTok | Output / MTok | Cache read / MTok | Cache write / MTok |
|---|---|---|---|---|
| Fable 5.2 | $8.00 | $40.00 | $0.20 | $10.00 |
| Opus 5.5 | $4.00 | $20.00 | $0.20 | $5.00 |
| Sonnet 5.5 | $2.00 | $10.00 | $0.10 | $2.50 |
| Haiku 5.5, prompt ≤100K tokens | $0.10 | $0.50 | $0.01 | $0.125 |
| Haiku 5.5, prompt >100K tokens | $0.50 | $2.50 | $0.05 | $0.625 |
Source: Anthropic pricing, checked November 2026. Prompt caching rates shown are for a 5-minute time-to-live (TTL); longer cache durations cost more.
The Claude context window can be up to 1 million tokens, depending on the model. Other charges may apply: web search costs $10 per 1,000 searches, plus token charges for processing the results; code execution costs $0.05 per container-hour after 50 free hours per day per organization; Managed Agents add $0.08 per active session-hour, plus token rates. US-only inference costs 1.1× standard input and output rates. Fast mode for Opus 5.5 and Fable 5.2 costs 2× standard rates for up to 2.5× speed.
Cost per task, not cost per token
A useful estimate needs both the input and output token counts. Use this formula:
(input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)
The examples below use the standard rates above, checked November 2026. They are token-cost estimates, not total project costs.
Customer support reply: 800 input tokens and 400 output tokens on Haiku 5.5, assuming the prompt is under 100K tokens:
- Input: 800 ÷ 1,000,000 × $0.10 = $0.00008
- Output: 400 ÷ 1,000,000 × $0.50 = $0.0002
- Estimated total: $0.00028
Code review of a 400-line file: 5,000 input tokens and 800 output tokens on Sonnet 5.5:
- Input: 5,000 ÷ 1,000,000 × $2 = $0.01
- Output: 800 ÷ 1,000,000 × $10 = $0.008
- Estimated total: $0.018

40-page contract summary: 60,000 input tokens and 1,500 output tokens on Opus 5.5:
- Input: 60,000 ÷ 1,000,000 × $4 = $0.24
- Output: 1,500 ÷ 1,000,000 × $20 = $0.03
- Estimated total: $0.27
The same contract on Haiku 5.5: At these token counts, the request is below Haiku’s 100K threshold:
- Input: 60,000 ÷ 1,000,000 × $0.10 = $0.006
- Output: 1,500 ÷ 1,000,000 × $0.50 = $0.00075
- Estimated total: $0.00675
At these rates, the Opus 5.5 example costs 40 times as much as the Haiku 5.5 example. That is a price comparison, not a claim that the models will produce equivalent results for every contract.
The three levers that cut an Anthropic bill
1. Cache repeated prompt content. A Sonnet 5.5 cache read costs $0.10 per MTok, 95% less than its $2 standard input rate. Cache writes cost more than standard input, so the savings come from reusing the cached prefix. It is most useful when requests repeatedly send the same large system prompt or context. It is usually not worth implementing for one-off prompts or prompts whose system content keeps changing.
2. Use batch processing when a response can wait. Batch processing reduces input and output token rates by 50% and runs asynchronously. It can fit tasks such as queued document summaries or offline classification, but not work that needs an immediate response.
Here is an illustrative monthly workload on Sonnet 5.5: 10,000 requests, each with a 5,000-token shared prefix, 1,000 other input tokens, and 500 output tokens. Assume the requests arrive in 100 bursts of 100 within each 5-minute cache window.
- Without caching or batch: 60 million input tokens × $2/MTok + 5 million output tokens × $10/MTok = $170 estimated.
- With batch processing alone: the same token volumes at half the rates = $85 estimated.
- With caching alone: 500,000 prefix tokens written × $2.50/MTok, 49.5 million prefix tokens read × $0.10/MTok, 10 million other input tokens × $2/MTok, and 5 million output tokens × $10/MTok = $76.20 estimated.
This comparison assumes every burst reuses the same prefix within the 5-minute TTL. Actual costs depend on request timing, cache hits, and token usage. Batch and prompt caching discounts can also be combined; these examples show each lever separately.
3. Route work to the right model. Start with a lower-cost model for tasks such as classification or extraction, then test whether it meets your quality requirements. Use a more capable model where the task needs it. Model choice can have a larger effect than small prompt edits.

The Haiku 5.5 threshold nobody notices
Haiku 5.5’s input rate rises from $0.10 to $0.50 per MTok and its output rate from $0.50 to $2.50 when a request’s prompt exceeds 100,000 tokens. The higher rates apply to that request, including when some input tokens are cache reads or writes. If you send long documents, check the full prompt size before estimating cost.
Legacy models and what they cost
These legacy rates are in USD per million tokens. The figures are standard input and output rates, checked November 2026 at Anthropic pricing.
| Model | Input / MTok | Output / MTok |
|---|---|---|
| Sonnet 4.6 / Sonnet 4.5 | $3 | $15 |
| Opus 5 / Opus 4.8 / 4.7 / 4.6 / 4.5 | $5 | $25 |
| Fable 5.1 | $10 | $50 |
| Fable 5 | $10 | $50 |
| Haiku 4.5 | $1 | $5 |
For new work, use current models unless you have a specific compatibility, evaluation, or migration reason to stay on a legacy model.
Anthropic API versus a Claude subscription
Claude app plans such as Pro, Max, and Ultra have usage limits; they are not metered API plans. The API is billed according to token and feature usage. Anthropic’s pricing page lists $100 in monthly API credits with Max and $200 with Ultra; those credits do not turn API use into unlimited usage. Team standard seats are $20 per seat per month billed annually or $25 monthly, while Enterprise seats are $20 per seat per month billed annually, with API usage billed separately at API rates.
Monitoring API spend
For developers, Token Wise offers optional API monitoring that estimates spend from outbound request metadata for api.anthropic.com: model, request and response size, status, and timestamp. It does not capture prompt text or authentication tokens. Monitoring is off by default, and no API key is required. A browser extension can only observe API calls made from the browser; applications that call the API server-side are not visible to it. Token Wise also shows live per-prompt cost estimates in the Claude web interface. See how it works. It supports ChatGPT, Claude, Perplexity, Gemini, Grok, Poe, OpenRouter, ChatHub, TypingMind, and You.com.

FAQ
How much does the Anthropic API cost per 1,000 tokens?
Divide the per-million rate by 1,000. For example, Sonnet 5.5 costs $0.002 per 1,000 input tokens and $0.01 per 1,000 output tokens at standard rates, checked November 2026. Your actual charge depends on the model, token counts, caching, and other features used.
Is Claude API cheaper than ChatGPT API?
It depends on the models and workload. As one comparable published pair, Sonnet 5.5 and OpenAI GPT-6.2 Sol are each listed at $2 per million input tokens and $10 per million output tokens on their official pricing pages, checked November 2026. The same rates do not mean the models are interchangeable; evaluate both on your task.
Does prompt caching really save money?
It can, when requests reuse the same prompt content. On Sonnet 5.5, a cache read is $0.10 per MTok versus $2 per MTok for standard input, but you also pay for cache writes. For one-off prompts or frequently changing prefixes, caching may not pay off.
Is there a free tier for the Anthropic API?
The Claude app has a free plan. The API is pay-as-you-go and billed by usage. Check Anthropic’s pricing page and Console for current account terms.
Estimate your workload using its expected input and output tokens, then compare model choice, caching, and batch processing against your latency needs. Before committing, verify current rates on Anthropic’s pricing page.