Claude API Pricing 2026: Opus, Sonnet, Haiku per MTok

The Claude API costs $2 per million input tokens and $10 per million output tokens on Sonnet 5.5, $4 and $20 on Opus 5.5, and $0.10 and $0.50 on Haiku 5.5 for prompts up to 100,000 tokens. The top model, Fable 5.1, costs five times Sonnet 5.5 on both input and output. Batch requests are half price, and cache hits on Sonnet 5.5 cost a twentieth of its input price. Rates read 19 October 2026.

Our take: the list prices are the least interesting part of the bill. On the current models, the multipliers (caching, Batch, US-only routing, fast mode) and the model you route each request to decide the monthly invoice more than the headline rates do, as the workload table further down shows. A team that routes well and caches its system prompt pays a fraction of what the headline table suggests.

Toolradar data: on the Anthropic API profile on Toolradar, the software directory we run, the API is listed as an alternative to 23 tools, joint #16 of 483 API tools, as of October 2026. Among those 483 API tools, 73% have a free plan, and the Claude API is not one of them: Anthropic gives new users a small amount of free credit and bills usage after that.

Every price below comes from Anthropic's API pricing documentation, which quotes all prices in USD.

How much does the Claude API cost per million tokens?

Model	Input	Output	Cache hit	5-minute cache write	Best fit
Claude Fable 5.1	$10	$50	$0.25	$12.50	The hardest reasoning and agent work, where errors cost more than tokens
Claude Opus 5.5	$4	$20	$0.20	$5	Complex coding and long agent loops
Claude Sonnet 5.5	$2	$10	$0.10	$2.50	Most production workloads
Claude Sonnet 5	Same as Sonnet 5.5	Same as Sonnet 5.5	$0.20	$2.50	Teams already pinned to the previous Sonnet
Claude Haiku 5.5, prompts up to 100K tokens	$0.10	$0.50	$0.01	$0.125	Classification, extraction, routing, high-volume chat
Claude Haiku 5.5, prompts over 100K tokens	$0.50	$2.50	$0.05	$0.625	Long-document jobs that still need the cheap model

All prices are per million tokens (MTok). Mythos 5.1, a limited-availability model for verified organizations, carries the same rates as Fable 5.1. Older models such as Opus 4.5 through Opus 5 sit at $5 in and $25 out, Sonnet 4.6 at $3 and $15, and Haiku 4.5 at $1 and $5.

Sonnet 5's launch price became permanent. The documentation says the Sonnet 5 rates announced as introductory through 31 August 2026 are now standard, and the planned move to $3 and $15 on 1 September will not happen. A budget built on the higher number can come down.

Which Claude model gives the best price per task?

Haiku 5.5 is the price outlier: it costs one-twentieth of Sonnet 5.5 on short prompts. For tagging, routing, extraction and support replies that fit under 100,000 prompt tokens, nothing else in the lineup comes close on cost. The catch is the threshold. A request whose prompt crosses 100,000 tokens, cache reads and writes included, is billed at five times the short-prompt rate for the whole request.

Sonnet 5.5 is the default for most apps. It costs half of Opus 5.5 on both input and output, and Anthropic's own guidance in the same document points Sonnet at most production workloads and Opus at the most complex reasoning.

Opus 5.5 earns its premium only where a cheaper model fails often enough to cost you retries or human review. Fable 5.1 sits at two and a half times Opus 5.5, which makes it a model to call for specific steps, not to run every request through.

One hidden factor changes the comparison with older models. Claude 4.7 and later use a newer tokenizer that, per the documentation, produces about 30% more tokens for the same text. By our arithmetic, that makes Sonnet 5.5 behave like roughly $2.60 and $13 per million against Sonnet 4.6's $3 and $15 on identical text: still cheaper, but by less than the sticker says.

How do prompt caching and Batch cut the bill?

Caching discounts the repeated part of a prompt, and Batch halves everything for work that can wait. They stack, so a pipeline that uses both pays a small fraction of the list rate on its input.

Lever	What it changes	Where it pays
5-minute cache write	1.25x the base input price	Pays off after one cache read
1-hour cache write	2x the base input price	Pays off after two cache reads
Cache hit	0.05x input on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1, 0.1x on other models	Long system prompts, RAG context, multi-turn chat
Batch API	50% off input and output	Evals, backfills, nightly enrichment

Cache hits on the 5.5 models are cheaper than on older ones. Most models read cache at a tenth of the input price, while Opus 5.5 and Sonnet 5.5 read it at a twentieth, so a chat app with a 20,000-token system prompt pays for that prefix almost nothing after the first call.

Batch rates sit exactly at half the standard ones: Sonnet 5.5 drops to $1 and $5 per million, and Haiku 5.5 to $0.05 and $0.25 on short prompts. The trade is latency: Batch is asynchronous, so it suits jobs nobody is waiting on.

Devshot sends a daily brief on the tools and bills developers act on; get it at devshot.co to hear about the next rate change first.

What do fast mode, US-only inference and tools add?

These modifiers push the bill up, and each one is easy to switch on without noticing.

Add-on	Price	Note
Fast mode, Opus 5.5	$8 input / $40 output per MTok	Double the standard Opus 5.5 rate; not available with Batch
US-only inference	1.1x on every token category	Applies to Claude 4.6 and later via inference_geo: "us"
Web search	1 cent per search, billed per 1,000	Plus the tokens of the results
Web fetch	No extra charge	You pay for the fetched tokens
Code execution	1,550 free hours a month, then $0.05 per container-hour	Free when used with the newer web search or web fetch tools
Managed Agents runtime	$0.08 per session-hour	Billed only while a session is running

US-only routing is the quiet one. A compliance team that pins inference_geo to the US pays 10% more on input, output and cache alike, for every request. Bedrock and Google Cloud price regional endpoints on their own terms, so the multiplier only describes Anthropic's first-party API and its marketplace versions.

Tool definitions also cost tokens. The tool-use system prompt adds 286 tokens per request on Opus 5.5, Sonnet 5.5 and Haiku 5.5 when tool choice is auto, and the browser-use toolset adds about 6,600 input tokens. Cache those definitions and the overhead nearly disappears.

What a real workload costs on each model

To make the multipliers concrete, here is one monthly workload priced on each current model: 100 million input tokens and 20 million output tokens, roughly a busy internal assistant or a mid-size support bot. The totals are our arithmetic on the published rates, and the tool, search and runtime charges above are left out.

Setup	Monthly total	What drives it
Fable 5.1, standard	$2,000	Top-tier rates on every token
Opus 5.5, standard	$800	Input and output each $400
Sonnet 5.5, standard	$400	Input and output each $200
Sonnet 5.5, Batch	$200	Half price, asynchronous
Sonnet 5.5, 80% of input from cache	$248 plus cache writes	$48 of input, $200 of output
Haiku 5.5, standard, short prompts	$20	Prompts kept under 100,000 tokens

The cached row assumes 80 million input tokens are cache hits at Sonnet 5.5's cache rate and 20 million are fresh input at the standard rate. Each cached prefix is written once at $2.50 per million for a 5-minute cache, a small add-on when the same prefix serves many calls.

Routing is the biggest lever on this table. Sending only the hard 10% of traffic to Opus 5.5 and the rest to Haiku 5.5 would cost about $98 a month on the same volume, against $800 for Opus on everything. The exact split depends on your evals, but the spread between models is wider than any discount Anthropic sells.

How does Claude API pricing compare with OpenAI?

The two price ladders now line up almost rung for rung. OpenAI's API pricing page lists gpt-6.1-sol at exactly Sonnet 5.5's input and output rates for short-context requests. Its gpt-6-astra matches Fable 5.1, and gpt-6-luna matches Haiku 5.5 on short prompts.

The difference is in the long prompts. OpenAI doubles input on long context (gpt-6.1-sol goes to $4 input and $15 output), while Anthropic says Claude 4.6 and later models, except Haiku 5.5, bill the full 1-million-token context window at the standard rate. For RAG over large documents, that makes Sonnet 5.5 the cheaper of the two mid-tier models; for short chat, it is a draw on the sticker, and tokenizer differences decide the real bill.

Our best LLM APIs ranking compares providers beyond these two, and the free LLM APIs guide covers the free tiers that Anthropic does not offer. Toolradar keeps a profile of the OpenAI API as well.

Is the Claude API cheaper than a Claude subscription?

It depends on volume, and the per-token meter is the reason. A subscription caps the monthly bill while the API bills every token, so heavy daily use by one developer tends to favour a plan, and a long agent session on Opus 5.5 can run up API cost that a flat plan would absorb. Our Claude Code pricing page prices the plans against API usage, the Claude Code review covers the tool itself, and how to use Claude Code walks through setup. Developers choosing an editor can read Claude Code vs Cursor and GitHub Copilot vs Claude Code.

The API wins for products, batch jobs and anything that runs without a person in the loop. Teams comparing assistants for code can read ChatGPT vs Claude for coding and the AI coding assistant pricing roundup, and those routing several providers through one endpoint should look at our API gateways list.

How we checked these prices

All Claude prices, multipliers, tool fees and the free-credit policy were read on 19 October 2026 on Anthropic's API pricing documentation. OpenAI's rates were read the same day on its API pricing page. The Toolradar figures come from its Anthropic API profile as of October 2026.

Workload totals, the routing example and the tokenizer adjustment are our arithmetic on those published rates, not quotes or measurements; we did not run the workload. Volume discounts and enterprise rates are negotiated with Anthropic's sales team and are not reflected here. Louis Corneloup, founder of Toolradar and Dupple, edited this page.

FAQ

How much does the Claude API cost in 2026?

Per million tokens, Claude Sonnet 5.5 costs $2 input and $10 output, Opus 5.5 costs $4 and $20, Haiku 5.5 costs $0.10 and $0.50 for prompts up to 100,000 tokens, and Fable 5.1 costs $10 and $50. Batch requests cost half, and cache hits on Sonnet 5.5 cost 5% of its input price.

Is there a free tier for the Claude API?

No ongoing free tier. Anthropic's pricing page says new users get a small amount of free credit to test the API, and enterprise teams can ask sales about extended trials. After that credit is used, usage is billed in USD, by card on standard accounts and by invoice for enterprise customers.

How much does Claude Opus 5.5 cost per million tokens?

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on the standard API. Batch halves both rates, cache hits cost $0.20 per million, and fast mode doubles them to $8 and $40 for quicker output.

What is the cheapest Claude model?

Claude Haiku 5.5, at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, or $0.05 and $0.25 through Batch. Prompts over 100,000 tokens move to five times those rates, so keep long-document jobs on Sonnet 5.5 or check which costs less for your prompt sizes.

Does Claude charge more for long context?

Not on Claude 4.6 and later models other than Haiku 5.5: Anthropic bills the full 1-million-token context window at the standard per-token rate, so a 900,000-token request costs the same per token as a short one. Haiku 5.5 is the exception, with higher rates once a prompt passes 100,000 tokens.

How much does prompt caching save on the Claude API?

A cache hit costs 5% of the input price on Opus 5.5 and Sonnet 5.5, 2.5% on Fable 5.1 and 10% on other models. Writing to the cache costs 1.25x the input price for a 5-minute cache or 2x for a 1-hour cache, so the 5-minute cache pays for itself after a single reuse.

Does US-only data residency cost extra?

Yes. Setting inference_geo to "us" on Claude 4.6 and later models applies a 1.1x multiplier to input, output and cache tokens on the first-party API and Claude Platform on AWS. Global routing, the default, uses standard prices.

Cite this: Devshot, "Claude API Pricing 2026: Opus, Sonnet, Haiku per MTok", November 2026.

Devshot is the daily brief for developers who watch what their stack costs. Get it at devshot.co.

Some offers on this page may be paid placements or contain affiliate links.