Claude API cost per request depends on the model's per-token price and on how many tokens each request uses. Claude Haiku 5.5 is the cheapest Claude model at $0.10 input and $0.50 output per million tokens, and it beats the 326-model median on every task I ran. Sonnet, Opus and Fable cost more per token, and Anthropic's tokenizer also produces about 39 percent more tokens than OpenAI's for the same English text.
Claude list prices
These are list prices per 1M tokens, as of 2026-10-11:
| Model | Input | Output | Cached input | Context | Reasoning mode | LMArena Text |
|---|---|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 | $0.50 | $0.01 | 1,000,000 | default | not listed |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.10 | 1,000,000 | always | 1476 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | 1,000,000 | always | 1507 |
| Claude Fable 5.1 | $10.00 | $50.00 | $0.25 | 1,000,000 | always | 1501 |
Sonnet costs 20 times Haiku on input and output. Opus is twice Sonnet, and Fable is five times Sonnet. Haiku and GPT-6 Luna share the same list price, so any cost gap between them on a task comes from token use, not the rate card.
Cost per request on four tasks
Haiku and Sonnet results are measured. Opus and Fable results are estimated. The last column is the median of all 326 models.
| Task | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | Fable 5.1 | Median, 326 models |
|---|---|---|---|---|---|
| Answer a chat question | 0.07¢ | 1.3¢ | 2.5¢ | 6.3¢ | 0.24¢ |
| Summarize a 20-page PDF | 0.22¢ | 4.5¢ | 8.9¢ | 22¢ | 0.83¢ |
| Review 500 lines of code | 0.66¢ | 8.0¢ | 16¢ | 40¢ | 1.4¢ |
| 50-step coding agent (modeled) | 6.6¢ | 84¢ | $1.68 | $3.46 | 48¢ |
Haiku is the only Claude model below the median on every row. On the PDF summary, Haiku costs $2.20 per 1,000 tasks and Sonnet costs $44.58, about 20 times as much for the same job. Here is the PDF task across popular models:
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GLM 5.3 FlashZ.ai | 0.10¢$0.99 / 1,000 | 1474 | 12,995 / 1,814 | measured best value | |
| 2 |  DeepSeek V4.1 FlashDeepSeek | 0.12¢$1.22 / 1,000 | 1475 | 12,970 / 1,992 | measured best value | |
| 3 |  GPT-6 LunaOpenAI | 0.16¢$1.57 / 1,000 | 1443 | 12,987 / 533 | measured | |
| 4 |  Claude Haiku 5.5Anthropic | 0.22¢$2.20 / 1,000 | - | 18,004 / 790 | measured | |
| 5 |  Qwen3.8 FlashAlibaba Qwen | 0.32¢$3.21 / 1,000 | - | 13,119 / 2,642 | measured | |
| 6 | GPT-5.4 NanoOpenAI | 0.33¢$3.26 / 1,000 | 1401 | 12,987 / 533 | estimated | |
| 7 |  Mistral Small 4Mistral | 0.33¢$3.31 / 1,000 | - | 13,373 / 2,170 | estimated | |
| 8 |  Llama 4 MaverickMeta | 0.42¢$4.16 / 1,000 | 1327 | 13,015 / 2,638 | estimated |
See the full task pages for chat answers, PDF summaries and code review.
The tokenizer tax
In my tokenizer comparison on the same English document, Anthropic models produced 1.39 times the tokens of OpenAI (OpenAI = 1.00). Other providers ranged from 0.98 to 1.07.
The PDF task shows it. Haiku 5.5 and Sonnet 5.5 each read 18,004 input tokens for the same 20-page PDF. GPT-6 Luna and GPT-6.1 Sol read 12,987. That is a ratio of about 1.39.
On document-heavy tasks, input is most of the bill, so 39 percent more input tokens raises cost even at an equal per-token price. Sonnet and GPT-6.1 Sol both list at $2.00 input and $10.00 output, yet the PDF summary costs 4.5¢ on Sonnet and 3.1¢ on GPT-6.1 Sol. Output tokens also differ (857 against 482), so the tokenizer is not the only factor.
The ranking inside the Claude family does not change. But when comparing Claude with OpenAI models, compare token counts as well as price per million.
What a 50-step agent costs
My modeled agent uses an 8,000-token system prompt, 50 steps, 1,500 more context tokens per step and 300 visible output tokens per step. Sonnet reads about 3.1 million input tokens and writes 20,790 output tokens per run.
| Model | With caching | Without caching | Per 1,000 runs (cached) |
|---|---|---|---|
| Haiku 5.5 | 6.6¢ | $0.34 | $66.35 |
| Sonnet 5.5 | 84¢ | $6.52 | $841.64 |
| Opus 5.5 | $1.68 | $13.04 | $1,683.28 |
| Fable 5.1 | $3.46 | $32.60 | $3,461.14 |
Caching cuts Sonnet from $6.52 to 84¢ and Fable from $32.60 to $3.46. If an agent repeats a long prompt on each step, set up caching before choosing a model. See the agent task page for more models.
Picking a model
- Haiku 5.5: short tasks, simple output, high volume. It was the cheapest Claude model on every task.
- Sonnet 5.5: a stronger model when per-task cost still matters. Its LMArena Text score is 1476, against 1507 for Opus and 1501 for Fable.
- Opus 5.5 and Fable 5.1: when quality matters more than cost. Fable is the most expensive Claude model on every task, at 6.3¢ for a chat answer and 40¢ for a code review.
To estimate your own volume, use the calculator. The methodology page explains how I count tokens and costs.
How I measured
Measured models ran once per task at default settings through OpenRouter, with output capped at 16,000 tokens. Estimated figures are labeled as such. Reasoning tokens are billed as output tokens. The PDF is 10,023 words of J. S. Mill's On Liberty, and the code review input is CPython's textwrap.py at 497 lines.
FAQ
Is Claude Haiku 5.5 cheaper than Sonnet 5.5 per request?
Yes. Haiku costs $0.10 input and $0.50 output per million tokens, against $2.00 and $10.00 for Sonnet. On the PDF summary, Haiku cost 0.22¢ per task and Sonnet 4.5¢.
Why does the same document cost more on Claude?
Anthropic's tokenizer produced 1.39 times the tokens of OpenAI's on the same English text. In the PDF task, Claude models read 18,004 input tokens where GPT-6 Luna read 12,987.
How much does caching lower Claude agent costs?
A lot, because cached input is priced far below regular input. On the 50-step agent, Sonnet fell from $6.52 per task without caching to 84¢ with it.