Anthropic released Claude Opus 5.5 today. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. That second number is the one that matters for most of the companies I work with.
The price cut is bigger than the headline
Input and output tokens dropped 20%, to $4 and $20 per million. The larger change is in cache reads, which dropped 60%, to $0.20 per million tokens.
Cache reads work like this. Most requests to Claude start with the same material every time: the system prompt, the tool definitions, a long reference document. Rather than reprocess that material on every call, Anthropic stores it after the first request and charges a discounted rate to read it back on subsequent ones. The more repetitive your requests, the more of your bill is cache reads. Agents with long tool lists that make dozens of calls per task are the extreme case; a one-shot document summarizer barely uses them.
That is why Anthropic's 40% figure is an average over typical workloads, not a guarantee. If your work is agentic, you should land near it or above. If it is not, expect something closer to the 20% token price cut plus whatever you save from Opus 5.5 using fewer tokens to finish the same task. Early testers reported real savings on that second point. Box saw it use a third of the tokens Opus 5 did, with answers 40% less verbose without losing accuracy.
Default effort is now the right setting
The thing I heard most from clients on Opus 5 was that the good results required cranking effort up, which cost time and money. That appears to have changed. Factory called Opus 5.5 the first model they would default to at medium effort; in their testing it matched Opus 5 on high effort while using 20 to 25% fewer output tokens. Deloitte reported that at its lowest effort setting, Opus 5.5 caught 72% of known bugs in code reviews compared to Opus 5's 56% at high effort. Medium is also the new default, so a request that sets no effort level now runs at medium where Opus 5 ran at high. If you have effort pinned to high in production, test medium before you assume you need it.
It writes better
Verbose, jargon-heavy output was the other consistent complaint about Opus 5. Anthropic addressed it directly. Opus 5.5 leads with what matters, uses less jargon and fewer odd turns of phrase, and sticks to the writing rules you give it. That last point is the practical one for me. Most of my engagements include a style file that tells Claude how a particular company writes. A model that actually follows those rules means less post-editing and fewer prompt workarounds.
Two integration changes to check before you switch
Opus 5.5 is no longer available with thinking mode switched off. If any of your pipelines pass a thinking-disabled parameter, they will need updating. It also no longer supports forced tool use: a request that forces Claude to call a particular tool returns an error. Pipelines that use that to get structured output, which is a common pattern for extraction work, need to move to strict tool use or structured outputs. The migration guide has the full list.
Anthropic also shipped an anti-distillation measure called preserved thinking, which rejects requests where the earlier part of a conversation, including the system prompt and the tool list, has been changed after Claude reasoned over it. It applies only to API accounts created on or after August 31, 2026, and unless your code edits earlier turns or swaps instructions or tools mid-conversation, you will never encounter it.
The safeguards will occasionally route you to a different model
Because of its capability in biology and cybersecurity, Anthropic is deploying Opus 5.5 with safeguards similar to those on Fable 5.1. In practice, finding and fixing bugs in your own code still works as normal, but Anthropic hands most other cybersecurity requests to Opus 4.8. For ordinary business and coding work this will not come up. If you are in life sciences, you can apply to a verification program now. For security work, Anthropic says it will open its cyber program to Opus 5.5 in the coming weeks.
What I'd actually do
For most clients: change the model string to claude-opus-5-5, re-run your eval set at medium effort, compare cost per task against Opus 5, and ship if the quality holds. Do not re-architect anything. Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks, with many of the same improvements, so hold off on re-planning your model tiers until the whole family is out.
Set your expectations correctly going in. The cost reduction is the certain part. The quality gain is likely smaller than the benchmark table implies: Anthropic itself says that at this level, benchmark margins are a less reliable guide to real-world differences, and that in its own use the gap between Opus 5.5 and Fable 5.1 is narrower than the scores suggest. The realistic outcome is Fable-level quality at Opus prices, not a step change in what your agents can do. That is still a good reason to switch.
Source: Anthropic, "Introducing Claude Opus 5.5" (September 22, 2026).