Claude API Pricing (November 2026): Cost per Million Tokens, per 100k Tokens and per Request

The Claude API costs, per million tokens in and out: $0.10 and $0.50 on Haiku 5.5 (prompts up to 100,000 tokens), $2 and $10 on Sonnet 5.5, $4 and $20 on Opus 5.5, and $10 and $50 on Fable 5.1, per Anthropic's pricing page, checked on 5 November 2026. So 100,000 input tokens cost between $0.01 and $1, and a typical request costs between $0.00035 and $0.035. The Batch API halves every rate, cache reads cost 10% of input or less, and US-only inference adds 10%. Below: the full list, the cost per 100k tokens and per request, the modifiers, and what one real-sized job costs on each model.

---

Claude API prices per million tokens

From Anthropic's pricing page, standard (global) rates, US dollars, checked 5 November 2026:

Claude Mythos 5.1 is priced like Fable 5.1 but is limited to organizations verified through Anthropic's programs. Opus 4.5 to 4.8 stay on the page at $5 and $25, Sonnet 4.5 at $3 and $15. Opus 4.1, Opus 4, Sonnet 4 and Haiku 3.5 are retired on the Claude API and remain only on some cloud platforms.

Two notes on the list. Sonnet 5's $2 and $10 is now permanent: it launched as introductory pricing, and the increase to $3 and $15 scheduled for 1 September 2026 "will not occur". And cache reads are cheaper on the newest models: 0.05x input on Sonnet 5.5 and Opus 5.5, 0.025x on Fable 5.1, against the usual 0.1x.

What 100k tokens cost on Claude

Most searches for Claude pricing are really one question: what does a given amount of text cost? Here it is for the four current models, standard rates:

Price per token is the per-million price divided by a million: one input token on Sonnet 5.5 costs $0.000002. Output costs five times input on every current model, so a job that writes long answers costs more than one that reads long documents.

How many tokens is $1? On input, 10 million on Haiku 5.5, 500,000 on Sonnet 5.5, 250,000 on Opus 5.5 and 100,000 on Fable 5.1. On output, a fifth of that.

One caveat on Haiku 5.5: a single request whose prompt goes over 100,000 tokens pays $0.50 and $2.50 for the whole request. So 100,000 tokens spread over many small requests cost $0.01; one 150,000-token prompt costs $0.075 in input.

What one request costs

Take a typical request: 2,000 tokens in (instructions, context and the question) and 300 tokens out.

An internal tool making 3,000 such requests a working day, about 66,000 a month, costs about $23 a month on Haiku 5.5 and about $460 on Sonnet 5.5. Agents cost more per task than this: each tool call is another request that resends the conversation, which is where caching pays.

Which Claude model is cheapest?

Claude Haiku 5.5, at $0.10 and $0.50 per million tokens, $0.05 and $0.25 with the Batch API. It is cheaper than every older Haiku and twenty times cheaper than Sonnet 5.5 on input.

Cheapest per token is not cheapest per job. A model that fails a task costs you a retry, a human check or a wrong answer. Use Haiku 5.5 for classification, extraction, routing and summaries where it passes your accuracy check; use Sonnet 5.5 or Opus 5.5 where it does not.

What changed with Haiku 5.5

Anthropic released Haiku 5.5 on 7 October, per launch coverage, as claude-haiku-5-5 on its API, AWS, Google Cloud and Microsoft Azure. Three things matter for cost:

• It is priced by prompt length. Up to 100,000 tokens: $0.10 and $0.50. Over 100,000 tokens, the whole request pays $0.50 and $2.50, cache reads and writes included. Sonnet and Opus bill the full 1M-token context at standard rates; Haiku 5.5 rewards short prompts.
• The per-token cut is 90%; the per-job cut is smaller. Claude 4.7 and later models use a newer tokenizer that Anthropic says "produces approximately 30% more tokens for the same text". The same text on Haiku 5.5 therefore costs about 87% less than on Haiku 4.5, not 90%. Coverage quotes Anthropic at about 75% cheaper on average, with its new adjustable reasoning counted. Reports differ; measure your own.
• It is now the default for volume work. Classification, extraction, routing, summaries, support triage: jobs where a small model passes the accuracy check now cost cents per thousand items.

The modifiers that move a bill

• Batch API: 50% off input and output, for requests that can wait for asynchronous processing. Not available with fast mode.
• Prompt caching: writing to the 5-minute cache costs 1.25x input, the 1-hour cache 2x. Reading costs 0.1x input (0.05x on Sonnet 5.5 and Opus 5.5, 0.025x on Fable 5.1). A 5-minute cache pays for itself after one read, a 1-hour cache after two. Caching stacks with the batch discount.
• US-only inference: 1.1x on every token category when you set inference_geo: "us" on Claude 4.6 and later models. Global routing is the default.
• Fast mode on Opus 5.5: $8 and $40, double the standard rate, for faster output. On Opus 5 and Opus 4.8 it is $10 and $50.
• Long context: Claude 4.6 and later models, except Haiku 5.5, bill the full 1M-token window at standard rates. A 900k-token request pays the same per token as a 9k-token one.
• Tools: web search costs $10 per 1,000 searches plus tokens. Web fetch has no extra charge beyond the tokens it brings in. Code execution is free alongside web search or web fetch; on its own, each organization gets 1,550 free hours a month, then $0.05 per container-hour. Any tool definition adds input tokens: 286 for the tool-use system prompt on the 5.5 models.

Claude on AWS, Google Cloud and Azure

• Claude Platform on AWS and Claude in Microsoft Foundry are run by Anthropic and bill the standard Claude API rates through the AWS or Azure marketplace, in units worth $0.01 each.
• Amazon Bedrock and Claude on Google Cloud (Google managed) are run by the cloud provider, which sets the price. Anthropic's page sends you to Bedrock pricing and Google Cloud pricing. For Claude 4.5 and later models, their regional and multi-region endpoints cost 10% more than global ones.

If your company already has cloud credits or a committed spend, running Claude through that cloud can be the cheapest route even at equal list prices.

Claude API vs a Claude subscription

The subscriptions are for people; the API is for software. Prices from claude.com/pricing on 5 November 2026:

If you are one person using Claude to write and code, a subscription is cheaper and has no per-token surprises. If you are putting Claude inside a product or a workflow, you need the API, and at the per-request costs above a small app often costs less than a single Pro seat. More on the plans in our Claude Team plan guide.

What a real workload costs

Take a common job: classify 100,000 support tickets, each with a 1,500-token prompt (instructions plus the ticket) and a 200-token answer. That is 150 million input tokens and 20 million output tokens, counted on the newer tokenizer that all the 5.x models share.

The cached column assumes 1,000 of the 1,500 input tokens are the same instructions on every call, read from cache, and leaves out the small cost of writing the cache. The gap between rows is a factor of 100. The gap between columns is a factor of two. Picking the model is the decision; the discounts are tuning.

That matches what we see on real projects. When we read an insurer's entire claims book, 13,000 documents, 96% of them photos, the model bill for the whole book came to under $25. The expensive parts were access to the documents, permissions and domain knowledge, never tokens.

How Claude compares with OpenAI

OpenAI moved the same month. GPT-6.1 Terra launched on 26 October at a reported $1 per million input tokens and $5 output, half the price of the models it replaces; we could not open OpenAI's own announcement to confirm. That is Haiku 4.5's price and ten times Haiku 5.5's short-prompt rate, though Terra is aimed at agent and computer-use work rather than bulk classification.

List prices are not the comparison that matters. Two models at different prices that both pass your accuracy check are a cost decision; a cheap model that fails it is not cheaper. Run the same 200 real items through each candidate, score them against answers you trust, then price the ones that pass. We wrote about that method in smaller LLMs, same results.

What we do, and what we don't do yet

What we run: our classification service calls models through the LiteLLM library (version 1.90.2) and OpenRouter, with the model chosen per task, so moving a step to a cheaper model is a configuration change, tested against the same scored sample. Most of the tasks configured in it today use Google models, picked on accuracy for each task.

What we sell: a $1,900 fixed-price starter that connects one workflow to your assistant or app, on your own data, with read and draft tools, in about two weeks, credited against a production build. Scoping a larger workflow costs $2,300 to $4,700, a first workflow in production starts at $17,500, and running one, models and hosting included, starts at about $470 a month.

What we don't do yet: we have not moved a production workload to Haiku 5.5. The price makes it the first candidate for every high-volume step, and the test is the same as for any model: our scored sample, before and after. We are not an Anthropic partner and earn nothing on your API spend.

Related

• How much an AI agent costs to build and run
• Cost efficiency for LLMs in production
• Claude Team plan: what it costs
• Anthropic's startup program