Claude Sonnet 5.5 costs $2 per million input tokens and runs faster
Key takeaways
- Sonnet 5.5 is priced at $2 per million input tokens and $10 per million output tokens.
- Anthropic says it is faster and more efficient than Sonnet 5, and pitches it at enterprise coding, document creation and spreadsheets.
- With enterprise reported at around 80 percent of Anthropic's business, price per token is becoming the main point of competition between labs.
Last updated: 5 October 2026
Two dollars in, ten dollars out. That is the reported price of Claude Sonnet 5.5 per million tokens, released on 28 September and described by Anthropic as faster and more efficient than Sonnet 5.
It is pitched at enterprise coding, document creation and spreadsheets, which is where most of the company's revenue is said to come from.
The numbers
Input tokens, the text you send, cost $2 per million. Output tokens, the text the model writes back, cost $10 per million. Output is five times the price of input, which is typical because generating text is the expensive part.
A quick example makes it concrete. A job that reads 500,000 tokens and writes 50,000 costs $1.00 for the input plus $0.50 for the output, so $1.50 in total. Very roughly, 500,000 tokens is a few hundred pages of text.
Why the price matters more than the name
According to the weekly roundup we used, enterprise accounts for around 80 percent of Anthropic's business. Companies at that scale run millions of requests, so a small change in price per token turns into a large change in their bill.
That is why we read the price as the headline. When several labs ship models of similar quality within weeks of each other, the cheaper and faster one is easier to defend in a budget meeting. Our piece on OpenAI's agents for enterprise and consumers shows the same push for business customers from the other side.
What it does not tell you
Price per token is not price per task. A faster model that needs fewer retries can cost less in practice, and a cheaper one that needs more can cost more. We have not tested Sonnet 5.5 ourselves, so treat the speed claim as Anthropic's.
If you would rather avoid per token pricing entirely, running a model on your own hardware is an option, covered in our guide to running a local LLM on a Mac.
What to watch
Watch whether rival labs respond with their own price cuts in the next few weeks. The quiet signal will be whether anyone starts advertising cost per finished task instead of cost per token.
