Claude API Pricing 2026: What Opus, Sonnet and Haiku Cost per Token and per Task
AIHub
Claude API pricing in 2026 runs from 0.10 dollars to 10 dollars per million input tokens and from 0.50 to 50 dollars per million output tokens, depending on the model. Claude Haiku 5.5 costs $0.10 in and $0.50 out, Claude Sonnet 5.5 $2 and $10, Claude Opus 5.5 $4 and $20, and Claude Fable 5.1 $10 and $50. Output always costs five times input. Prompt caching cuts repeated input to between 2.5 and 10 percent of the normal rate, and the Batch API halves everything.
Those are list prices from Anthropic’s own pricing page, read on 2026-11-02. Per-token numbers are hard to budget with, so this guide also prices six real jobs on each model, from a single question to a 30-step coding agent task, and shows where the API beats a 20 dollar Claude Pro subscription and where it does not.
Claude API prices per million tokens
| Model | Input | Output | Cache write (5 min) | Cache read | Batch input / output |
|---|---|---|---|---|---|
| Claude Fable 5.1 | $10 | $50 | $12.50 | $0.25 | $5 / $25 |
| Claude Opus 5.5 | $4 | $20 | $5 | $0.20 | $2 / $10 |
| Claude Sonnet 5.5 | $2 | $10 | $2.50 | $0.10 | $1 / $5 |
| Claude Haiku 5.5, prompts up to 100K tokens | $0.10 | $0.50 | $0.125 | $0.01 | $0.05 / $0.25 |
| Claude Haiku 5.5, prompts over 100K tokens | $0.50 | $2.50 | $0.625 | $0.05 | $0.25 / $1.25 |
US dollars per million tokens, standard global routing, from Anthropic’s pricing documentation on 2026-11-02. A one-hour cache write costs twice the input price instead of 1.25 times.
Older models are still on sale and some cost more than their replacements. Claude Opus 4.1 and Opus 4 are $15 and $75, almost four times Opus 5.5. Opus 5, 4.8, 4.7, 4.6 and 4.5 are $5 and $25. Sonnet 4.6, 4.5 and 4 are $3 and $15, while Sonnet 5 matches Sonnet 5.5 at $2 and $10 (its cache reads are $0.20 rather than $0.10). Claude Haiku 4.5 is $1 and $5, ten times Haiku 5.5. If an application is pinned to an old model ID, moving it to the current generation is usually a price cut as well as an upgrade.
Two families sit outside the main lineup. Claude Mythos 5.1 is priced like Fable 5.1, and Claude Fable 5 and Mythos 5 are also $10 and $50 but with cache reads at $1.
What common jobs cost on each model
The token counts below are the ones used across this site’s guides, so the numbers line up with ChatGPT alternative for writing and ChatGPT alternative for coding. No caching unless stated.
| Job (tokens in / out) | Haiku 5.5 | Sonnet 5.5 | Opus 5.5 | Fable 5.1 |
|---|---|---|---|---|
| Short question (1,000 / 500) | $0.0004 | $0.007 | $0.014 | $0.035 |
| Ten-turn conversation (39,000 / 5,000) | $0.006 | $0.13 | $0.26 | $0.64 |
| 1,500 word article with one edit (4,000 / 4,000) | $0.002 | $0.05 | $0.10 | $0.24 |
| Summary of a 100-page PDF (65,000 / 1,000) | $0.007 | $0.14 | $0.28 | $0.70 |
| 30-step coding agent task (600,000 / 30,000) | $0.08 | $1.50 | $3.00 | $7.50 |
| Same agent task, 80% of input cached | $0.03 | $0.59 | $1.18 | $2.82 |
The ten-turn conversation costs far more than ten short questions because every turn resends the earlier ones: 300-token prompts and 500-token answers add up to 39,000 input tokens by the end. The agent task is the same effect at scale. An agent rereads the files and the conversation on every step, so input dominates, and that is exactly the input that caching discounts. The cached row counts only cache reads; the first write of each prefix adds a little on top at 1.25 times the input rate.
Two patterns stand out. Haiku 5.5 is cheap enough that cost stops being a reason not to use it: a thousand short questions cost about 35 cents. And the gap between Sonnet and Opus is exactly 2x on every job, so the choice between them is about quality per task, not about a price cliff.
The modifiers that change the bill
Prompt caching. If you send the same system prompt, document or conversation history repeatedly, mark it for caching. A cache read costs a tenth of the input price on most models, and less on the newest: 5 percent on Opus 5.5 and Sonnet 5.5, and 2.5 percent on Fable 5.1 and Mythos 5.1. A five-minute cache write costs 1.25 times input, so it pays for itself on the first reuse; a one-hour write costs 2 times and pays off on the second.
Batch API. Requests that can wait (overnight classification, bulk summaries, evaluation runs) go through the Batch API at half price on input and output. Batch and caching stack. Summarising 1,000 PDFs on Sonnet 5.5 is about 140 dollars at standard rates and about 70 through batch.
Fast mode. Claude Opus 5.5 has a fast mode at $8 and $40, double the standard price, for faster output. It does not combine with batch.
Data residency. Pinning inference to the US with inference_geo: "us" multiplies every token price by 1.1 on Claude 4.6 and later models. The default global routing is standard price.
Long prompts. Claude 4.6 and later models, except Haiku 5.5, bill a 900,000-token request at the same per-token rate as a 9,000-token one, across the full 1M context window. Haiku 5.5 is the exception: past 100,000 prompt tokens the whole request moves to $0.50 and $2.50. Even then a 150,000-token prompt costs 7.5 cents on Haiku 5.5 against 30 cents on Sonnet 5.5.
Tools. Web search is 10 dollars per 1,000 searches plus the tokens of the results. Web fetch costs only the tokens it pulls in. Code execution is free when used with web search or fetch; on its own each organisation gets 1,550 free container hours a month and pays 5 cents an hour after that. Defining tools also adds a few hundred input tokens of system prompt to every request (286 on the current models).
Managed Agents. Sessions on Claude Managed Agents pay normal token prices plus 0.08 dollars per hour of running time. Anthropic’s own example, a one-hour Opus 5.5 coding session with 50,000 tokens in and 15,000 out, comes to 58 cents.
The tokenizer catch
A per-token price is only comparable if the token counts are. Anthropic says Claude 4.7 and later models use a newer tokenizer that produces about 30 percent more tokens for the same text than earlier Claude models. So Sonnet 5.5 at $2 per million tokens is a smaller saving over Sonnet 4.5 at $3 than the list prices suggest: the same prompt is roughly 1.3 times as many tokens. Across vendors the effect is unknown in advance, which is why the AI model API pricing data lists tokenizer differences under what it does not measure. Before budgeting a large workload, run a sample of your real prompts through Anthropic’s token counting endpoint and multiply from there.
Claude against OpenAI, Google and the market
| Model | Input | Output |
|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 |
| GPT-6 Sol | $2 | $10 |
| Gemini 3.1 Pro | $2 | $12 |
| Grok 4.7 | $2 | $6 |
| Claude Haiku 5.5 | $0.10 | $0.50 |
| GPT-6 Luna | $0.10 | $0.50 |
| Gemini 3.5 Flash Lite | $0.30 | $2.50 |
Vendor list prices per million tokens, from the AI model API pricing data checked 2026-11-02. At the frontier the big three have converged on 2 dollars in, and Haiku 5.5 now matches OpenAI’s smallest current model to the cent. Against the market as a whole Claude is still at the expensive end: across 335 paid models on OpenRouter the median is $0.40 in and $1.88 out, and Anthropic has the highest median of any large provider, because its catalog is mostly frontier and premium models. The full distribution, by provider and context window, is in the LLM API pricing statistics.
Claude API or Claude Pro?
Claude Pro is 20 dollars a month, or 17 on the annual plan, and includes Claude Code. Using the job table: 20 dollars at API rates buys about 2,850 short Sonnet 5.5 questions, about 155 ten-turn conversations, or about 13 uncached 30-step agent tasks (about 34 with caching). If you have fewer than five real conversations a day, or code with an agent a few times a week, paying per token is cheaper. If you are in Claude all day, the subscription wins, and when Pro’s session limit runs out, its usage credits bill at the same API rates as the table.
The plan side of that comparison, including how Pro’s five-hour sessions work and what ChatGPT Plus gives for the same 20 dollars, is in Claude Pro vs ChatGPT Plus. If you would rather pay per message without managing API keys, AI chat aggregators explain the four ways those apps bill. AIHub is one of them, so read that with the disclosure in mind: it charges Claude Opus 5.5, Sonnet 5 and Haiku 4.5 at Anthropic’s list price in credits, alongside GPT and Gemini, on one bill.
How to keep a Claude API bill small
- Default to Sonnet 5.5 and move a task to Opus only when Sonnet’s answer is measurably worse on that task. The price gap is exactly 2x.
- Send classification, extraction and routing to Haiku 5.5. At a tenth of a cent per thousand short calls, cost stops mattering.
- Cache every prefix you resend: system prompts, tool definitions, reference documents, conversation history in agents.
- Push anything that can wait a few hours through the Batch API.
- Cap output with
max_tokenson jobs that do not need long answers. Output is five times the input price on every Claude model. - Check the model IDs in production. An app still calling Opus 4.1 pays 15 and 75 dollars for work Opus 5.5 does at 4 and 20.
Prices change. Every figure here is dated 2026-11-02; check Anthropic’s pricing page before signing a contract or launching a feature.
