What is Claude Opus 5.5, released on 22 September 2026?
Anthropic released Claude Opus 5.5 on 22 September 2026, under the model ID claude-opus-5-5. It is the first model of the 5.5 family: Sonnet 5.5 and Haiku 5.5 will follow “in the coming weeks”, according to the announcement. For a European company already running Claude in production, one question comes quickly: should you switch, and for which use cases?
Claude Opus 5.5 is Anthropic’s new flagship model, presented as performing at the level of Claude Fable 5.1 on most tasks while costing 40% less than Opus 5 on typical workloads. It no longer runs with thinking mode switched off, applies preserved thinking to recent API accounts and ships with watermarking designed for the EU AI Act.
What Anthropic announced, and what it left open
On performance, Anthropic says Opus 5.5 “performs at the level of Claude Fable 5.1 on most work”. On cost, it announces a 40% reduction compared with Opus 5 “on typical workloads”, and output generated more than 30% faster. The model is available on the Claude Platform, in Claude Code and on the major clouds: Amazon Web Services, Google Cloud and Microsoft Azure. Zero data retention remains available, as with previous Opus models.
The list price per million tokens falls by 20% on both input and output, and by 60% for cache reads. Anthropic attributes the 40% to that lower price per token combined with a model that uses fewer tokens per task, and that second part inevitably varies between workloads: an agentic chain that often rereads the same cached context will not save in the same proportion as a one-off extraction. Several points also remain open at the time of writing: no retirement date has been set for Opus 5, and the definition of a “typical workload” is not published. The real saving depends on your own request profile, and you have to measure it yourself.
What does it change for a European enterprise?
The release affects budget, user experience and compliance, with very different weight depending on the use case. A CIO running a dozen Claude use cases will not see the same benefit everywhere, and that unevenness should decide the order of migration.
The cost of high-volume use cases
Use cases that call Opus thousands of times a day, such as document review, contract analysis or tier-2 support agents, are where a reduction of this size actually moves the budget. On an assistant consulted a few dozen times a week, the difference is too small to justify a project.
An application that stays on Opus 5 will capture none of this reduction. The gain ships under a new model ID, and someone has to go and get it. We described this mechanism in our analysis of AI model lock-in and the architecture that lets you switch: a price cut on an existing model ID reaches applications with no effort, whereas a new generation requires a migration.
Latency in interactive use cases
Output that is more than 30% faster matters most where a person is waiting for the answer: a legal assistant used during a meeting, a copilot for a bank adviser, an agent that chains steps and whose calls add up. For overnight batch processing, latency changes nothing about the service delivered.
Teams that had ruled Opus out because of response times can reopen the question, especially for agents that drive screens through computer use or call APIs, where each step carries its own delay and the user’s wait accumulates.
Watermarking for the EU AI Act
Anthropic states that Opus 5.5, like Fable 5.1, comes with its watermarking measures designed to comply with the EU AI Act. This relates to Article 50 of the regulation, which sets transparency rules for AI-generated content and has applied since 2 August 2026. For the marking obligation, the Digital Omnibus only granted a grace period, until 2 December 2026, to generative systems placed on the market before 2 August 2026. Opus 5.5, released on 22 September, falls outside that window: its marking is expected from launch.
Watermarking on the provider side leaves the deploying company’s own obligations intact, starting with informing people exposed to generated content. Our briefing on the EU AI Act timeline after the delay sets out what is still due, and our page on EU AI Act compliance for enterprises describes how we support it.
Should you migrate to Claude Opus 5.5 right away?
Not all at once. Migrate high-volume or latency-sensitive use cases first, after replaying them on an evaluation set built from your real data. For use cases currently running on Sonnet or Haiku, wait for the 5.5 versions announced for the coming weeks before deciding.
The order we recommend to our clients:
- List the use cases that call Opus today, with their volume and criticality.
- Check when each API account was created and how preserved thinking affects your integrations.
- Adapt requests to the parameters the new model rejects.
- Replay a sample of real requests on Opus 5.5 and compare quality, consumption and latency.
- Switch a first use case through configuration, with a way back.
- Reassess Sonnet and Haiku use cases once the 5.5 versions are out.
When switching quickly makes sense
The natural candidates are use cases already on Opus 5 where cost is holding back scale, and those where response time hurts the user experience. Then there are use cases built on Fable 5.1 for its quality: if Opus 5.5 holds the announced level, it is the most direct cost-saving route for them, provided your evaluations confirm it on your own data.
When waiting is the better call
Use cases running on Sonnet or Haiku follow a different calculation. Moving from Sonnet to Opus, even a cheaper Opus, is still a change of tier, and Sonnet 5.5 is on its way. Applications recently signed off after a long acceptance cycle, particularly in regulated sectors where every model change goes back to a committee, are better off bundling the migration with the rest of the family than running two acceptance cycles within a few weeks.
What does preserved thinking change for your integrations?
Preserved thinking is an anti-distillation safeguard introduced with Fable 5.1. According to Anthropic, it stops API users from editing Claude’s prior context in an attempt to extract its reasoning. It applies to Fable 5.1 and Opus 5.5 for API accounts created on or after 31 August 2026, and Anthropic points to documentation explaining how to test and update affected integrations.
An API account opened on or after 31 August therefore behaves differently from an older one. A new project, a new subsidiary or an integrator taking over a programme under a fresh account all fall into this case. Integrations that modify the conversation history between calls are the ones to test first, since that is exactly what the safeguard governs.
Thinking mode becomes mandatory
Anthropic states that Opus 5.5 is no longer available with thinking mode switched off. For use cases that called Opus without extended reasoning, such as field extraction or document classification, behaviour changes: the model reasons before answering, and those thinking tokens are billed as output tokens even when their text is not returned. Consumption per request and response time change accordingly. The announcement refers to an adjustable effort level, set to “medium” by default in its comparisons, which becomes the main lever for these simpler use cases. One more reason to measure your own workloads instead of transposing the 40% figure.
How do you test Opus 5.5 before switching production?
A model change should go through acceptance testing like any software update, with real data and success criteria written down before the test. Teams that decide on the strength of a good demo tend to find regressions in production, where they cost the most.
Replay your real cases rather than benchmarks
The benchmarks Anthropic publishes measure generic tasks. Your contract review, complaint triage or support agent has its own traps: in-house formats, domain vocabulary, edge cases the team took months to identify. An evaluation set of a few hundred anonymised production requests, each with its expected answer, is usually enough to decide. For teams already scaling Claude Code across the enterprise, that set belongs in continuous integration: a candidate model that fails the suite stays out of production.
Measure at the use-case level
The overall monthly bill says nothing about the effect of a model change. Consumption, latency, failure rate and quality need to be attributed to each use case, so that Opus 5 and Opus 5.5 can be compared on the same scope. Without that measurement, the announced reduction remains a vendor promise that nobody can confirm or challenge in a budget review. This tracking is part of running agents in production in our Claude Ops agent industrialisation programme.
Adapt requests before switching
Changing the model ID is not enough. Anthropic’s migration guide lists settings that claude-opus-5-5 rejects with an error: thinking disabled or a manual thinking budget, forced tool calls, modified temperature or sampling parameters, a prefilled response and, on the Claude API and Google Cloud, the older computer use tool version. Code that reads the response by position also breaks, since a response can now start with thinking blocks, and tool loops must pass those blocks back unmodified. These changes happen in code, before any switch, and they go through the evaluation set like everything else.
Plan the rollback
Once requests are compatible, switching a use case comes down to a configuration setting. If indicators deteriorate after go-live, the team moves back to Opus 5 without redeploying the application. As long as Anthropic has not announced a retirement date for Opus 5, that safety net is available, and it is better to have planned it before you need it.
Who should decide on a model change?
According to Deloitte (State of AI in the Enterprise, 2026 edition), only 21% of organisations have a mature governance model for their AI agents. Elsewhere, the model in production is often whatever the team that shipped the first prototype picked, and that choice has never been revisited. A release like Opus 5.5 exposes the gap: nobody has the mandate to say whether the gain is worth an acceptance cycle.
We recommend naming an owner for the model choice of each use case, with written criteria: a quality threshold on the evaluation set, a cost ceiling, compliance requirements, hosting constraints. In the LOOP™ AI governance methodology, a model change is treated as a change to the agent itself, with a recorded approval and a rollback criterion.
When hosting is imposed
Many large organisations consume Claude through their cloud provider, for contractual, data residency or security reasons. Opus 5.5 is announced on Amazon Web Services, Google Cloud and Microsoft Azure, but regional availability and retention terms can differ from one platform to another. Checking these points in your region before acceptance testing avoids approving a model your contract does not yet let you call.
What should you take away for your AI roadmap?
Across a portfolio of Claude use cases, the Opus 5.5 release translates into a fairly short work list: find the Opus calls, fix incompatible requests, replay the evaluation set, switch. Organisations with an abstraction layer, an evaluation set and a named owner can capture the announced reduction on their heaviest workloads this quarter. For the rest, every new generation opens a project, the decision slips to the next budget cycle, and Sonnet 5.5 then Haiku 5.5 will land before the first question is settled.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. A cheaper model addresses part of the first problem, provided you go and get it. To see where your organisation stands, the AI maturity assessment scores you on the 6 axes of the Koneetiv framework (2026 edition). And if you are looking for a Claude integration partner to run this kind of migration, Koneetiv, an official Anthropic partner and Claude pure player, supports large organisations with their Claude agents in production.