Claude Opus 5.5: The API Details That Matter Before You Migrate

Claude Opus 5.5 is more than a new model name to paste into an existing request. The important question is whether an integration still matches the model’s documented request and response behavior.

Anthropic’s model overview lists the API model ID as claude-opus-5-5, with a release date of September 22, 2026. It also lists a 1-million-token context window and a maximum standard output of 128,000 tokens. The documented API price is $4 per million input tokens and $20 per million output tokens.

The migration risk is in the controls, not the version number

Opus 5.5 uses adaptive thinking, and the documentation says thinking is always enabled. Manual thinking configurations that disable thinking or set a manual thinking budget are rejected. Instead, output_config.effort controls the model’s reasoning depth, with medium listed as the default in the effort documentation.

That changes how max_tokens should be interpreted. Thinking tokens count as output tokens and also count toward max_tokens, so the limit covers more than the visible prose returned to a reader. A request that worked with an older thinking configuration may therefore need both its thinking settings and token ceiling reviewed.

The useful takeaway is that the model swap is a configuration change, not just an identifier change. The version string is the easy part; the control surface is where compatibility is decided.

Tool use has a separate compatibility boundary

The Opus 5.5 changes page documents tool_choice: auto as supported, while forced choices using any or a named tool are rejected. That means an application designed around mandatory tool selection should not assume that replacing the model ID will preserve the same request behavior.

The same documentation says that assistant thinking and redacted_thinking blocks must be preserved unchanged when a tool-use loop sends the conversation back to the API. The thinking guide also explains why an interface can appear quiet during a long operation: progress between tool calls may arrive as thinking blocks rather than ordinary text blocks.

These are two different failure modes. A rejected tool-choice setting is an API compatibility problem. A client that renders only text blocks is a display problem. Treating both as “the model stopped responding” makes diagnosis harder.

The price is a token rate, not a task price

The pricing documentation lists separate input and output rates: $4 per million input tokens and $20 per million output tokens. Because thinking tokens are counted as output tokens, visible answer length alone is not enough to understand usage.

The practical comparison is therefore not “Is Opus 5.5 expensive per task?” There is no single task price in the documentation. A meaningful evaluation should compare the workload’s input tokens, output tokens, effort setting, tool behavior, and resulting quality. That produces a more useful decision than comparing model labels or a single sample response.

A focused preflight for an existing integration

Before switching a request to claude-opus-5-5, check four things:

1. Use the documented model ID and choose an output_config.effort value deliberately; medium is the documented default.
2. Remove incompatible manual thinking settings and make sure max_tokens leaves room for both thinking and visible output.
3. Use the documented supported tool-choice behavior rather than assuming forced tool selection still works.
4. Preserve thinking blocks through tool-use turns and render thinking blocks when the interface needs to show progress.

The broader point is simple: Opus 5.5 should be evaluated as an API contract update with a reasoning control, not as a label change. The model overview tells you what is available; the thinking, effort, and tool-use documentation tells you whether an existing integration is ready for it.