Running a bot on Claude now costs 20% less
Anthropic released Claude Opus 5.5 on 22 September, and the price came down with it. A million units of input cost five dollars before and now cost four. A million units of output cost 25 dollars before and now cost 20. One billing unit is called a token, and a million tokens run to roughly 1,500 pages of text.
A flat monthly subscription does not get cheaper, but it does get bigger. Anthropic raised the usage limits on the Pro, Max, Team and Enterprise plans. Anyone paying by usage, through the programming interface or through one of the cloud providers, sees the cut land straight on the monthly bill.
The context window holds a million tokens, and the model takes images as well as text, so scanned receipts and invoices can go in directly. Its knowledge stops at June 2026, so anything that happened after that has to come from you.
The new price against the old
| Opus 5 | Opus 5.5 | |
|---|---|---|
| Million input tokens | $5 | $4 |
| Million output tokens | $25 | $20 |
| Cache read | $0.50 | $0.20 |
| Cache write | $6.25 | $5 |
Requests sent as a batch rather than in real time cost half of those rates, two dollars for input and ten for output.
Repeated text is where the money is
The row that moved most is the third one, down 60%. A cache is where the provider keeps text you have already sent, so it does not have to process it from scratch on every call.
That matters to a business more than it sounds. A bot answering customers sends the same instructions, the same price list and the same product catalogue on every single request. In an automation that runs for a long stretch, most of the text passing through is repeat text. On a bill like that, the drop from 50 cents to 20 is worth more than the 20% off the headline rate.
Treat 40% as a direction, not a budget line
Anthropic says a typical run of the new model costs about 40% less than a run of Opus 5, and that it works at the level of Claude Fable 5.1 on most jobs. Read that number carefully, because the comparison sits across different effort settings: Opus 5.5 at medium effort matches or beats Opus 5 at high effort on coding tests, while the scores quoted in headlines were measured at maximum effort.
The scores themselves are real and reported. On Terminal-Bench 4.0 the model scored 66.4% against 55.8% for Fable 5.1. Output comes out more than 30% faster. Anyone building a monthly budget should run a week on their own real load and read the bill, rather than multiplying last month by 0.6.
If someone already built you an automation
Four things changed at the programming-interface level, and an automation someone built for you on Opus 5 may stop working after the switch. The model's thinking step can no longer be turned off. The option to force it to use a particular tool is gone. Thinking blocks are now tied to conversation state. Older computer-use tools get rejected.
Opus 5 is not going away. Anthropic marks it as active and commits to keeping it in service until at least 22 September 2027, which leaves a full year to update code that leans on it. Claude Sonnet 5.5 and Claude Haiku 5.5, the cheaper models, are due in the coming weeks with the same efficiency work behind them.