Moving from Claude Opus 5 to Opus 5.5: what changes

Anthropic released Claude Opus 5.5 on 22 Sep 2026. If your firm runs software that calls Claude Opus 5, the new model is cheaper per token and finishes most work in fewer of them, so the question is when to move and what has to change first. This post covers what breaks, what changes quietly, what it costs and how to run the switch without surprising your users. It is written for the person accountable for the integration, with the parameter names included for whoever picks up the ticket.

You do not have to move yet

There is no deadline. Anthropic's model deprecations page still lists Claude Opus 5 as Active, with a retirement date of "not sooner than July 24, 2027". Anthropic's models overview has moved it to its list of legacy models, which remain available. Claude Opus 5.5 carries a retirement date of not sooner than 22 Sep 2027. Anthropic commits to at least 60 days' notice before it retires a publicly released model, and Amazon Bedrock and Google Cloud set their own schedules separately.

So a firm on Opus 5 has ten months before the earliest possible retirement. The reason to move sooner is money and speed. The reason to wait a little is that the upgrade changes how requests are written, and some of the open-source tooling between your code and Anthropic was still catching up in launch week.

The upgrade is a change you make

Nothing switches by itself. Every integration names the model it wants, and yours currently says claude-opus-5. Moving means changing that to claude-opus-5-5 (on Amazon Bedrock the ID is anthropic.claude-opus-5-5). Each integration moves on its own, and it is normal to move them at different times. If you use Claude Managed Agents, Anthropic's migration guide says the model name is the only change required. For everything else, it is where the work starts.

Four things that will fail outright

Anthropic documents four breaking changes for code running on Opus 5. Each produces an error, so they are the easy ones to find in testing.

1. Thinking can no longer be switched off. Opus 5.5 always reasons before it answers. A request that sets thinking to disabled, or that sets a manual thinking budget, is rejected with a 400 error. Integrations that turned thinking off to save time or money need to remove that setting and use the effort parameter instead, which is now the only control over how much the model reasons.
2. Forced tool use is rejected. Code that told the model it must call a tool (the tool_choice types any and tool) gets a 400 error. Anthropic's replacement is to leave tool choice on auto, mark tools as strict for schema-valid output, and tell the model in the prompt when the tool applies. Pipelines that forced a tool call to get structured data back need rethinking rather than a flag change.
3. Thinking blocks are tied to the model and the conversation. Each reasoning block records which model produced it. Opus 5.5 reads blocks from Opus 5 and earlier models, so a live conversation can move up to it and keep its reasoning, but moving away from Opus 5.5 drops that reasoning silently. And if your code edits the system prompt, the tool list or an earlier message part-way through a conversation, replaying a reasoning block after the edit returns a 400 error on accounts created on or after 31 Aug 2026. Keeping conversations append-only avoids the question.
4. The older computer-use tool is rejected on the Claude API and Google Cloud. Integrations that drive a screen using the computer_20251124 tool must move to the computer_toolset_20260801 toolset. On Amazon Bedrock the older tool continues to work.

Anthropic's what's new page has the exact error text and the before-and-after request for each.

One change that fails without an error

The fifth change is the one most likely to reach a customer. On Opus 5, the short notes the model writes between tool calls ("checking the order history", "found three matching records") came back as ordinary text, and applications that stream them to a user as progress updates displayed them. On Opus 5.5 those notes come back inside thinking blocks, whose text is empty at the default display setting. No request fails. The application simply goes quiet while the model works.

If you have an agent or assistant that shows people what it is doing, test this specifically. The fix is to set thinking.display to a value that returns the text (updates, in beta, for the progress notes alone, or summarized for the notes plus a summary of the reasoning) and render the blocks that come back.

The setting that changes without an error or a message

The effort parameter controls how hard the model works on a request, and therefore how much it costs and how long it takes. Opus 5 defaulted to high. Opus 5.5 defaults to medium. A request that never set effort explicitly now runs one level lower than it did, with no error and no warning.

Anthropic's position is that the levels do not mean the same thing across models: in its own testing, Opus 5.5 at medium matched or exceeded Opus 5 at high on coding and knowledge work. That may hold for your workload, but it is Anthropic's measurement on Anthropic's tests. Its prompting guide says to set effort explicitly and re-run your own evaluations at several levels rather than carry the old setting over. Two related points: at a given level the new model thinks more per turn than Opus 5, so a max_tokens limit sized for Opus 5 with thinking off can now cut replies short, and changing effort between requests invalidates the prompt cache, so pick a level per workload and hold it.

What it costs

Anthropic publishes its per-token prices on its pricing page. Relative to Opus 5, Opus 5.5 is a fifth cheaper on both input and output tokens, and the Batch API halves those rates again. Cache reads (re-using a prompt prefix you have sent before) are priced at a twentieth of the input rate, where most models use a tenth, so a cache read costs 60% less than on Opus 5. For an agent that sends the same system prompt and tool definitions on every step, cache reads can be a large share of the bill, so that is often where the bigger saving sits. Our post on controlling Claude API costs covers caching, batching and effort in more detail.

Anthropic also says that at default settings Opus 5.5 "will cost 40% less than Opus 5 on typical workloads". That is a different claim from the price cut. It combines the lower rate with the model using fewer tokens and tool calls per task, measured on Anthropic's workloads at Anthropic's default effort. Budget on the list price and treat the rest as an upside to measure.

Where it runs and where your data goes

Opus 5.5 is available on the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry from launch, so whichever route your contract uses will have it. Anthropic has said Claude Sonnet 5.5 and Claude Haiku 5.5 "will follow in the coming weeks", without a date, so if your integration also uses those models they are separate decisions for later.

Data residency has not changed with this model. On the direct Claude API the only inference choices are global routing or US-only, and US-only carries a 10% price multiplier. There is no UK or EU option on the direct API. Amazon Bedrock's model card for Opus 5.5 lists an EU geo inference profile that "keeps data within EU regions", with London among the regions it serves, but no UK-only profile. A firm whose contracts require UK or EU processing needs a cloud route, and should read that route's residency terms rather than infer anything from the model name.

Tooling that was behind in launch week

Many firms reach Anthropic through an open-source layer such as LiteLLM or LangChain, or their engineers use Claude Code. All three had gaps on launch day. We re-checked each on 23 Sep 2026.

• Claude Code. The stable release channel was on version 2.1.267, and the model requires 2.1.280 or newer. As of 23 Sep the npm stable tag still pointed at 2.1.267, so anyone on the stable channel needs to move to the latest channel to use the model (issue 96130, open).
• LiteLLM. Support for the direct Claude API route was merged on 22 Sep, with the Bedrock, Google Cloud and Azure routes tracked as follow-up work in an open issue. Check that the version you run includes the change before relying on it.
• LangChain. An open issue from 23 Sep reports that ChatAnthropic can still send the forced tool choice and disabled-thinking settings that Opus 5.5 rejects, and that a partial fix covered structured output only.

These are launch-week reports and will date quickly. The lasting point is that a green test suite against Opus 5 says nothing about whether your framework version handles Opus 5.5, so pin versions deliberately and re-run the suite against the new model before production traffic reaches it.

How we would run it

This is the order we would follow. Inventory every place the model name appears, including middleware and configuration outside the code. Fix the four hard failures in a development environment and run the existing test suite. Then capture a sample of real traffic and score the model's output at two effort levels, medium and low, against the current Opus 5 baseline, adding high only if quality drops. Check every user-facing surface that shows progress while the model works. Set effort and max_tokens explicitly so the next model change cannot move them silently. Move one integration first, watch cost and quality for a week, then move the rest.

Anthropic ships a /claude-api migrate command in Claude Code that applies the model swap and parameter changes across a code base and produces a checklist to verify by hand. It is a good first pass. It does not run your evaluation or check your progress display, which is where the judgement sits.

If you are weighing whether to stay on Claude at all, our living comparison of current models has the cross-vendor view. For firms that would rather someone else carried the migration, including the evaluation and the tooling checks, that is work we do as part of Claude implementation. Request a Quote and a Claude-specialist engineer will scope it with you.

Frequently asked questions

Will my Claude integration switch to Opus 5.5 automatically?

No. Every integration names the model it calls, and yours keeps calling Claude Opus 5 until someone changes the model ID to claude-opus-5-5 and deploys the change.

When does Claude Opus 5 stop working?

Anthropic lists Claude Opus 5 as an active model with a retirement date of not sooner than 24 Jul 2027, and commits to at least 60 days' notice before retiring a model. Amazon Bedrock and Google Cloud set their own schedules.

Is Opus 5.5 really 40% cheaper?

The per-token price is 20% lower than Opus 5, and cache reads are 60% lower. The 40% figure is Anthropic's estimate of total cost on its typical workloads at default settings, which includes the model using fewer tokens per task. Your saving depends on your traffic and effort setting.

Does Opus 5.5 change where my data is processed?

No. On the direct Claude API the choices remain global routing or US-only inference. UK or EU processing is available only through cloud routes such as Amazon Bedrock's EU geo profile, and there is no UK-only option.