September 22, 2026
Long agent projects will cost less: Claude Opus 5.5 is up to 40% cheaper than Opus 5
On September 22, Anthropic released Claude Opus 5.5 at $4 per 1 million input tokens. The model charges $20 per 1 million output tokens, while Opus 5 cost $5 for input and $25 for output. Its 1 million-token context window fits an average novel in full.
Dan Shipper
@danshipper
BREAKING: Anthropic just released Opus 5.5, and it is already pulling some of our recent shifts to Codex back to Claude. We tested it at @every on coding, design, writing, and knowledge-work tasks. In our tests, it sometimes outperforms Fable 5.1 and costs up to 40% less than Opus 5. It is making a case to be the new everyday model: It is excellent at building complete projects and matching your taste. On large end-to-end projects, I have started choosing Opus over Fable more often. And @kieranklaassen is replacing Fable 5.1 with Opus in day-to-day product work. He calls it his favorite model. It writes clear text, but still trails Astra on writing tasks. Opus 5.5 scores 68.42 on readability, the best result among the models we have tested. But in my writing tests, it repeatedly buries the key point in opening paragraphs and edits. You will be able to understand what the model is trying to say. Hooray. But for everyday writing, it is still behind. The economics stand out here. Anthropic says Opus 5.5 will cost $5 per million input tokens and $20 per million output tokens. The input price is the same as Opus 5, while output is 20% cheaper. Overall, it should cut costs by about 40% versus Opus 5. It still has a habit of doing too much. @hammermt tested it on our standard knowledge-work benchmarks, and the results were excellent when it did return. In at least one test, it ran past the 10-minute limit before producing an answer. Bottom line: If you build apps and interfaces, try it. It has become my default model for ambitious coding projects. In daily work, I am still roughly 80/20 Codex over Claude, and I still prefer Sol and Astra for editing. But I am spending far more tokens with Claude than I was a week ago. Here is the picture: Codex still suits me better as a working setup, but Anthropic is catching up fast. They have made smaller models stronger than larger ones before, just look at Sonnet 3.7. It looks like the same happened with Opus 5.5. I will post the full vibe check at @every soon!
· 1.4M views
Previously, Opus 5 cost $5 per 1 million input tokens and $25 per 1 million output tokens. Now Opus 5.5 costs $4 and $20, while cache read has fallen from $0,50 to $0,20. Batch API offers a 50% discount on input and output tokens. To migrate a project in Claude Code, Anthropic suggests the command `/claude-api migrate this project to claude-opus-5-5`: it changes the model ID and required parameters, then provides a manual review checklist. In the API, replace `claude-opus-5` with `claude-opus-5-5` and set `output_config.effort` instead of the previous thinking settings, or the API will return 400.
Long sessions. Anthropic tailored the model for agents working over long periods on code and knowledge tasks. The maximum response is 128 thousand tokens, and reaches 300 thousand in the Batch API beta. Anthropic says the default settings reduce costs by 40% through cheaper tokens, cache read, and generation that is more than 30% faster. In the company's test, Opus 5.5 reviewed and fixed 200 thousand lines of code in under 3 hours, while Opus 5 took more than 20 hours and used 2,5 times as many tokens. Dan Shipper writes that some recent Codex users have already returned to Claude for daily large projects.
Independent runs will show whether the claimed savings match token usage and speed in working repositories.
