Source briefing. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers and teams putting AI into practical use
⚡ TL;DR — the 30-second version
- Use the exact ID claude-opus-5-5 and review thinking settings.
- Cache reads are $0.20/M; input/output prices are $4/$20 per million.
- Record safeguards and fallback routing in evaluation logs.
- A future Sonnet or Haiku announcement is not present availability.
Claude Opus 5.5 is available as claude-opus-5-5. Anthropic's model catalog lists a 1M context window, 128K maximum output and always-on adaptive thinking. The release notes state that disabled and manually enabled thinking configurations return an error; review the migration guide rather than carrying request options over unchanged.
Pricing has several moving parts. Standard input/output rates are $4/$20 per million tokens, while cache reads are $0.20/M. Anthropic reports lower token use per task as well as lower unit prices, leading to an estimated 40% reduction on its typical workloads. That percentage is a vendor result—not a guarantee for a support bot, code-review agent or document pipeline.
“The least interesting migration is replacing a string. The useful migration tests behavior, billing and failure modes.”
Preserved thinking is another compatibility consideration. The release describes anti-distillation controls for applicable accounts, and the documentation explains how modifications to prior context are handled. Test replayed conversations, middleware that rewrites histories, compaction, and any system that strips or edits reasoning blocks.
The production safeguards can route some sensitive tasks to other models. Anthropic says its reported Opus 5.5 evaluations were run with safeguards enabled and documents the fallback behavior. For your own evals, record the served route and refusals alongside correctness, rather than treating every response as an identical model run.
Start with a fixed regression set, shadow a small volume of traffic, record output-token usage and cache hits, and only then expand. Keep rollback explicit. Sonnet 5 remains current until a successor is actually available; the announcement that Sonnet 5.5 will follow is not an API release.
Field guide / 01
Create a compatibility matrix before changing traffic
Inventory each caller, not only the primary chat interface. Background jobs, evaluation scripts and repair loops may set different thinking options or rewrite conversation history. List their request fields, tool definitions, context transformations and expected error handling.
Run a small synthetic conversation through every caller. Include a tool failure, a long answer, a resumed thread and a request that triggers your own approval gate. Preserve the old configuration until the new path is observable.
- → Remove unsupported thinking configurations only after reviewing the documented migration behavior.
- → Test middleware that edits or compacts prior messages.
- → Record the model actually serving a safeguarded or fallback response.
Field guide / 02
Measure a complete multi-turn job
Cheaper cache reads can materially change an agent's cost, but a low input rate is not your entire invoice. Track ordinary input, cache writes, cache reads and all billable output across every turn. Keep review time and failed runs in the denominator.
Do not apply the provider's typical-workload savings percentage to your budget without measuring reuse. A mostly single-turn workload and a long cached coding session can respond very differently to the same price change.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Can I carry all Opus 5 request settings over unchanged?
Do not assume so. The release documents always-on thinking and preserved-context constraints. Validate every integration path against the current migration guidance.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Introducing Claude Opus 5.5Anthropic↗
- 02Claude Platform release notesAnthropic↗
- 03Claude models overviewAnthropic↗
- 04Claude Opus 5.5 system cardAnthropic↗
