Anthropic released Claude Opus 5.5 on September 22, opening the Claude 5.5 family with a model it says approaches Fable 5.1 on most work. The release combines lower token prices with changes that matter to developers who manage an agent’s conversation history themselves. Anthropic’s announcement
For an existing Opus integration, the first question is whether a normal session survives the migration intact. A successful first prompt does not exercise the history edits, tool changes and resumed sessions that happen later.
The price reduction has several parts
Anthropic lists these prices per million tokens:
| Token category | Opus 5.5 | Opus 5 |
|---|---|---|
| Input | $4 | $5 |
| Output | $20 | $25 |
| Cache reads | $0.20 | $0.50 |
| Cache writes | $5 | $6.25 |
The company reports roughly 40% lower cost on typical workloads at default settings, reflecting both prices and token use. That is a workload result, not a uniform discount on every request. Fast mode has separate input/output prices of $8/$40 per million. Pricing and efficiency details
An application that repeatedly reads a stable conversation will have a different cost profile from one that sends mostly fresh material. Compare actual input, output and cache usage before projecting a monthly saving. We have not benchmarked Opus 5.5 on S5 Labs workloads.
Custom agents need to check preserved thinking
Anthropic’s documentation says prior thinking blocks are tied to the context before them: the system prompt, tool definitions and earlier messages. Changing that prefix can invalidate the blocks. Prefix enforcement is on by default for API accounts created from August 31, 2026; older accounts can opt in. Preserved-thinking documentation
This reaches ordinary application behavior. Rebuilding a system prompt with a new date, shortening an old tool response, or changing the tool list can affect a continuing session. Anthropic recommends keeping the prefix stable and appending messages. Its documentation says Claude Code, claude.ai, Managed Agents and the Claude Agent SDK already handle request construction. Integration guidance
Model switching deserves a separate test. Opus 5.5 can read earlier Opus thinking, but cannot read Fable or Mythos thinking. Unreadable blocks are dropped for that request. A router can therefore complete a request while giving the selected model less prior reasoning than expected. Model-switching behavior
For a custom agent, save a representative multi-turn session and replay it through the new model. Include a restart and whichever history transformations the application actually performs. Check the resulting output as well as HTTP success: error-free execution alone will not reveal every loss of useful context.
Availability and the first evaluation
Anthropic lists availability across its platform, AWS, Google Cloud and Microsoft Azure, with claude-opus-5-5 as the Claude Platform identifier. Sonnet 5.5 and Haiku 5.5 are planned for the following weeks. Release availability
Start with a completed task whose result a reviewer can judge: a bug fix with a known reproduction, a document analysis with checkable citations, or a migration with acceptance tests. Record cost and completion time alongside the corrections the reviewer still needs to make. That provides a useful comparison with the current model without turning a launch claim into an assumed production outcome.
Our coverage of agent oversight metrics explains why review delays belong in that comparison too.