What happened
- Anthropic released Claude Opus 5.5 on September 22, 2026. According to the official model page, it costs 40% less to run than Opus 5 on typical workloads.
- The list price is $4 per million input tokens and $20 per million output tokens. Cache reads drop to $0.20.
- The company reports 66.4% on Terminal-Bench 4.0 versus 52.3% for Opus 5, and text generation more than 30% faster.
- The same page states that cybersecurity tasks are routed to Opus 4.8 and that work in biology requires prior approval through a verification program.
Why it matters
- For a team that bills in pesos and pays for inference in dollars, 40% less moves the threshold beyond which automating is cheaper than buying hours of work. That calculation changes today, not in the next budget cycle.
- The savings come with a condition the announcement doesn’t highlight: anyone who built code review or log analysis on Opus 5 will get an older model for those tasks without changing a line of their integration.
- The anti-distillation measure called preserved thinking applies only to accounts created after August 31, 2026. Old and new accounts don’t operate under the same rules.
The number
40% lower operating cost compared with Opus 5, according to Anthropic. It’s the figure the company uses to justify both the price cut and the increase in usage limits.
Context
Anthropic spent the last few weeks explaining why it was better to go slower: it gave outside auditors their own desks and acknowledged four incidents in which Claude got into real systems. The September 22 announcement points the opposite way: more capability, faster and cheaper, with expanded usage limits.
What’s next
- Sonnet 5.5 and Haiku 5.5 were announced for “the coming weeks,” with no committed date.
- The model is already available on Anthropic’s own platform, AWS, Google Cloud and Azure under the identifier
claude-opus-5-5. - The page includes the content marking required by the European AI Act, with no announced review date.
Bottom line
Two weeks ago the conversation was about how fast it makes sense to release models. Today the answer came from the price: 40% less, on the same day OpenAI cut its own in half.
Sources
Edited by Rodrigo Cornejo. How we select and verify each fact is in who writes.
