
Claude API pricing: what every Claude model costs per token, and the cheapest way to run it
Anthropic API pricing per 1M input and output tokens for Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2, with cache rates and real Messages API costs.
Claude API pricing on UnificAlly, per 1M input / output tokens, as of November 16, 2026: $0.015 / $0.075 for Claude Haiku 5.5 on prompts up to 100,000 tokens, $0.30 / $1.50 for Sonnet 5.5, $0.60 / $3 for the Claude Opus 5.5 API and $2.40 / $12 for Fable 5.2. Every Claude model bills input and output tokens separately, plus cache writes and cache reads. Anthropic released Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.2 between September 22 and November 9, so the model you priced in the summer is probably no longer the one to buy.
TL;DR: Start on Claude Opus 5.5. It ranks #4 on Arena's text board (1507, pulled November 16, 2026), and on our rate card it costs less per token than Opus 5, 4.8, 4.7 and 4.6. Swap Haiku 4.5 for Haiku 5.5: lower rate, 1M context instead of 200K, and a slightly higher Arena score. Then budget for the costs the rate card does not show: the newer tokenizer turns the same text into roughly 30% more tokens, and every request that carries tools pays 286 to 675 hidden system-prompt tokens depending on the model. All Claude models answer on
/v1/messages,/v1/chat/completionsand/v1/responseswith one key.
Claude API pricing per model (November 2026)
Anthropic API pricing, the first-party list, is on Anthropic's pricing page. Our Claude pricing per million tokens, with the cache lines that most comparison tables leave out:
| Model | Option | Our rate |
|---|---|---|
| Claude Fable 5.2 | input | $2.40/1M tokens |
| output | $12/1M tokens | |
| cache read | $0.06/1M tokens | |
| cache creation | $3/1M tokens | |
| Claude Fable 5.1 | input | $3/1M tokens |
| output | $15/1M tokens | |
| cache read | $0.30/1M tokens | |
| cache creation | $3.75/1M tokens | |
| Claude Fable 5 | input | $3/1M tokens |
| output | $15/1M tokens | |
| cache read | $0.30/1M tokens | |
| cache creation | $3.75/1M tokens | |
| Claude Opus 5.5 | input | $0.60/1M tokens |
| output | $3/1M tokens | |
| cache read | $0.03/1M tokens | |
| cache creation | $0.75/1M tokens | |
| Claude Opus 5 | input | $0.75/1M tokens |
| output | $3.75/1M tokens | |
| cache read | $0.075/1M tokens | |
| cache creation | $0.9375/1M tokens | |
| Claude Opus 4.8 | input | $0.75/1M tokens |
| output | $3.75/1M tokens | |
| cache read | $0.075/1M tokens | |
| cache creation | $0.9375/1M tokens | |
| Claude Opus 4.7 | input | $0.75/1M tokens |
| output | $3.75/1M tokens | |
| cache read | $0.075/1M tokens | |
| cache creation | $0.9375/1M tokens | |
| Claude Opus 4.6 | input | $0.75/1M tokens |
| output | $3.75/1M tokens | |
| cache read | $0.075/1M tokens | |
| cache creation | $0.9375/1M tokens | |
| Claude Sonnet 5.5 | input | $0.30/1M tokens |
| output | $1.50/1M tokens | |
| cache read | $0.03/1M tokens | |
| cache creation | $0.375/1M tokens | |
| Claude Sonnet 5 | input | $0.30/1M tokens |
| output | $1.50/1M tokens | |
| cache read | $0.03/1M tokens | |
| cache creation | $0.375/1M tokens | |
| Claude Sonnet 4.6 | input | $0.45/1M tokens |
| output | $2.25/1M tokens | |
| cache read | $0.045/1M tokens | |
| cache creation | $0.5625/1M tokens | |
| Claude Haiku 5.5 | input (prompt up to 100k) | $0.015/1M tokens |
| output (prompt up to 100k) | $0.075/1M tokens | |
| cache read (prompt up to 100k) | $0.0015/1M tokens | |
| cache creation (prompt up to 100k) | $0.01875/1M tokens | |
| input (prompt over 100k) | $0.075/1M tokens | |
| output (prompt over 100k) | $0.375/1M tokens | |
| cache read (prompt over 100k) | $0.0075/1M tokens | |
| cache creation (prompt over 100k) | $0.09375/1M tokens | |
| Claude Haiku 4.5 | input | $0.15/1M tokens |
| output | $0.75/1M tokens | |
| cache read | $0.015/1M tokens | |
| cache creation | $0.1875/1M tokens |
Haiku 5.5 is the only Claude model with two rows. Once a request's prompt passes 100,000 tokens, the whole request, output included, moves to the higher row, and Anthropic counts cached tokens toward the 100,000. Every other model in the table bills one rate whatever the prompt length, so a 900,000-token Opus 5.5 request costs the same per token as a 9,000-token one.
Each model has its own page and playground: Claude Fable 5.2, Fable 5.1, Fable 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Opus 4.5, Claude Sonnet 5.5, Sonnet 5, Sonnet 4.6, Sonnet 4.5 and Haiku 4.5.
Two lifecycle dates touch this list. Anthropic retires Sonnet 4.5 on November 30, 2026 and names Sonnet 5.5 as the replacement. Opus 4.5 is listed as retiring no sooner than November 24, 2026. Sonnet 5's launch price, first sold as introductory, is now Anthropic's standard price, and the increase planned for September 1 was cancelled.
Which Claude model to pay for: Opus 5.5 first, Haiku 5.5 for volume
Anthropic's own docs point new projects at Opus 5.5 and keep Fable 5.2 for demanding reasoning and long agent runs. The numbers back that up.
| Model | Arena text score (rank) | Context | Our input / output per 1M |
|---|---|---|---|
| Claude Opus 5.5 (high) | 1507 (#4) | 1M | $0.60 / $3 |
| Claude Fable 5.1 (max) | 1502 (#7) | 1M | $3 / $15 |
| Claude Opus 5 (high) | 1490 (#16) | 1M | $0.75 / $3.75 |
| Claude Sonnet 5.5 (xhigh) | 1473 (#47) | 1M | $0.30 / $1.50 |
| Claude Haiku 5.5 (high) | 1421 (#133) | 1M | $0.015 / $0.075 |
| Claude Haiku 4.5 | 1414 (#149) | 200K | $0.15 / $0.75 |
Scores come from Arena's text leaderboard, board dated November 11, 2026. Fable 5.2 was not on the board yet.
- Claude Opus 5.5 is the default. It sits three places above Fable 5.1 on Arena and costs less per token than any older Opus on our card. Effort defaults to
mediumon this model, one step below Opus 5, so setoutput_config.effortyourself if you are comparing the two. - Claude Fable 5.2 is for the hard tail. Anthropic released it on November 9, and our docs note it runs up to 30% faster than Fable 5.1. On our card it also costs less per token than Fable 5.1 and Fable 5, so nobody should start new work on either. Before paying for it, see how it fared against OpenAI's newest model in our GPT 6.2 Astra vs Claude Fable 5.2 test: identical grades, with Fable billing 6.5 times as much per graded call.
- Claude Sonnet 5.5 is the volume tier for coding and documents. It bills at the same rate as Sonnet 5 on our card, with the same 1M context. It trails Opus 5.5 by 34 Arena points, which is the gap you pay to skip.
- Claude Haiku 5.5 replaces Haiku 4.5. It costs less, holds five times the context and scores a little higher. Stay on Haiku 4.5 only if a pipeline was tuned against it and you have not re-tested.
That makes the cheapest Claude API in November 2026 Haiku 5.5 on prompts under 100,000 tokens, and the cheap Claude API choice for real coding work Sonnet 5.5. For Claude Opus pricing against Claude Sonnet pricing, compare the current pair: Opus API pricing starts at $0.60 per million input tokens on Opus 5.5, and Sonnet API pricing at $0.30 on Sonnet 5.5.
Claude token cost beyond the rate card
The rate card prices a token. Three rules decide how many tokens you buy, and none of them shows up in a price table.
The tokenizer changed. Claude Opus 4.7 and every model after it, Haiku 5.5 included, split text into roughly 30% more tokens than Sonnet 4.6 and older models did. Anthropic puts 1M tokens at roughly 555,000 English words on the new tokenizer, against about 750,000 on the old one. A classification job that ran 3M input tokens a month on Haiku 4.5 becomes about 3.9M tokens on Haiku 5.5: $0.45 on Haiku 4.5 against $0.0585 on Haiku 5.5. The move still saves money, just less than the per-token rates suggest.
Tools add a hidden system prompt. When a request carries a tools array, Anthropic injects its own tool-use instructions and bills them as input. Per request, with tool_choice set to auto or none:
| Model | Hidden tool-use tokens per request |
|---|---|
| Claude Opus 4.7 | 675 |
| Claude Opus 4.6, Sonnet 4.6 | 497 |
| Claude Haiku 4.5 | 496 |
| Claude Sonnet 5 | 354 |
| Claude Opus 4.8 | 290 |
| Claude Opus 5.5, Opus 5, Sonnet 5.5, Haiku 5.5 | 286 |
Your tool names, descriptions, JSON schemas, tool_use blocks and tool_result blocks are billed on top of that. An agent that makes 10,000 tool-carrying calls a month pays for 6.75M hidden tokens on Opus 4.7 ($5.0625) and 2.86M on Opus 5.5 ($1.716), before it sends a single word of its own.
Reasoning bills as output. Thinking is always on for Opus 5.5 and every Fable model, and the thinking tokens bill at the output rate even when the response hides them (the default display is omitted). Lower effort is the only dial on Opus 5.5, because thinking: {type: "disabled"} returns a 400 there.
Claude cache read and write pricing, and when caching does nothing
Anthropic charges 1.25x the base input rate to write a 5-minute cache entry and 2x for a 1-hour entry. A cache hit costs 0.1x on most models, 0.05x on Opus 5.5 and Sonnet 5.5, and 0.025x on Fable 5.2 and 5.1. So the 5-minute entry is already ahead on its first hit, while the 1-hour entry needs two hits to break even. On our side the same split shows up as the cache creation and cache read lines in the table above, and the Messages response reports them as cache_creation_input_tokens and cache_read_input_tokens.
The rule that catches people is the minimum cacheable prompt. Anything shorter caches silently as nothing:
| Minimum prefix | Models |
|---|---|
| 512 tokens | Fable 5.2, 5.1, 5, Opus 5.5, Opus 5, Sonnet 5.5, Haiku 5.5 |
| 1,024 tokens | Opus 4.8, Sonnet 5, Sonnet 4.6 |
| 2,048 tokens | Opus 4.7 |
| 4,096 tokens | Opus 4.6, Opus 4.5, Haiku 4.5 |
A 3,000-token system prompt caches on Haiku 5.5 and Opus 5.5. On Haiku 4.5 and Opus 4.6 it never caches, so every turn pays full input price. If your chatbot's cache_read_input_tokens stays at zero on an older model, check this table before you debug anything else.
Batch jobs are the other discount Anthropic offers first-party: the Message Batches API takes 50% off input and output, and cache rates apply on top. The rates in this guide are UnificAlly's standard per-request rates.
Claude API cost examples for real Messages API requests
The Claude API cost of a task is its token count times the rate. Worked at UnificAlly's Claude API price for each current model:
| Task | Tokens (in / out) | Input cost | Output cost |
|---|---|---|---|
| One support reply, Sonnet 5.5 | 1,000 / 500 | $0.0003 | $0.00075 |
| Summarize a 50-page PDF, Opus 5.5 | 30,000 / 2,000 | $0.018 | $0.006 |
| Route 10,000 tickets, Haiku 5.5 | 3M / 200K | $0.045 | $0.015 |
| A month of chat, Sonnet 5.5 | 20M / 4M | $6 | $6 |
The ticket job keeps each prompt far below 100,000 tokens, so it stays on Haiku 5.5's lower row. Feed Haiku 5.5 a 120,000-token contract and that single request bills on the higher row, input and output both. The chat month is where caching earns its keep: with a fixed system prompt over the 512-token minimum, most of those 20M input tokens can bill as cache reads instead.
Older pricing questions still come up, so for reference: Claude Opus 5 pricing on our card is $0.75 / $3.75, the same as Claude Opus 4.8 pricing and Claude Opus 4.6 pricing. Claude Sonnet 5 pricing matches Sonnet 5.5, and Claude Sonnet 4.6 pricing is $0.45 / $2.25. Claude Haiku 4.5 pricing (the Claude Haiku 4.5 API pricing people still search for) is $0.15 / $0.75. Claude Fable 5 pricing sits at $3 / $15, the same Claude Fable 5 price as Fable 5.1. Each figure is a Claude cost per 1M tokens, input then output.
Calling the Claude API on the Messages endpoint
Two ways in. Anthropic's Console gives you a key for api.anthropic.com at Anthropic's rates. A UnificAlly key covers every Claude model plus GPT, Gemini and the image, video and audio models, billed per token from one balance with no plan to buy.
Send LLM traffic to https://direct-api.unifically.com. Our docs warn that the main api.unifically.com host sits behind a proxy whose timeout can cut a long Claude answer short, and the tokens generated before the cut are still billed. An Anthropic-format request to Opus 5.5:
curl -X POST https://direct-api.unifically.com/v1/messages \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-d '{
"model": "anthropic/claude-opus-5-5",
"max_tokens": 2048,
"messages": [
{"role": "user", "content": "Explain prompt caching in two sentences."}
]
}'
With thinking on, thinking blocks can arrive before the text block, so read the block whose type is text instead of content[0]. The usage object carries input_tokens and output_tokens, plus the two cache fields when caching applied, and the response id doubles as the task id in GET /v1/tasks. The Anthropic SDK works with base_url set to https://direct-api.unifically.com. OpenAI SDK code can send the same model ID to /v1/chat/completions with base_url set to https://direct-api.unifically.com/v1.
Claude FAQ
How much does the Claude API cost?
On UnificAlly, per million input / output tokens: $0.60 / $3 for Opus 5.5, $0.30 / $1.50 for Sonnet 5.5, $0.015 / $0.075 for Haiku 5.5 under 100,000 prompt tokens, and $2.40 / $12 for Fable 5.2. A 1,000-token question with a 500-token answer on Sonnet 5.5 costs a fraction of a cent.
Is the Claude API free?
No. Anthropic and UnificAlly both bill per token. A new UnificAlly account starts with a $0.50 balance, which covers about 33 million Haiku 5.5 input tokens, and money you add has no expiry date.
Which Claude model is the cheapest?
Claude Haiku 5.5, at $0.015 input and $0.075 output per million tokens on UnificAlly for prompts up to 100,000 tokens. Longer prompts bill the whole request on its higher row. It replaced Haiku 4.5 as the cheapest Claude model in October 2026.
What does Claude Opus 5 cost per million tokens?
Claude Opus 5 costs $0.75 input and $3.75 output per million tokens on UnificAlly, with cache reads at $0.075. Its successor Opus 5.5 is $0.60 / $3.
Why is Anthropic Claude API pricing different from what I pay on UnificAlly?
Anthropic's page is the first-party rate card for api.anthropic.com. UnificAlly is an independent API with its own per-token rates, listed in the table above and billed from a single balance. The request shapes match, so moving code over means changing the base URL, the key and the model ID.
Do cached tokens cost less?
Yes. A cache read on Sonnet 5.5 costs $0.03 per million tokens on UnificAlly against $0.30 for fresh input. Writing the cache costs more than plain input, and prompts under the model's minimum, 512 tokens on the 5.5 models and 4,096 on Haiku 4.5, never cache.
Is Claude Fable 5 the same as Claude Mythos 5?
They are the same model sold two ways. Fable 5 is the generally available version, with Anthropic's safety classifiers on. Mythos 5 has the same list price and limits but is open only to organizations verified through Anthropic programs such as its Cyber Verification Program. API buyers get Fable 5, now succeeded by Fable 5.2.
To set Claude against the other two big labs, read the OpenAI API pricing guide and the Gemini API pricing guide, or try the Claude Opus 5.5 API and Claude Haiku 5.5 API in the playground.


