To use Opus 5.5 in Claude Code, first check your client version and provider, select the model explicitly, and confirm the session configuration before giving it work. A successful model switch is the setup milestone; a reviewed, working patch is the useful result.
Checked September 30, 2026. This is a documentation-based setup guide, not an Agent.Space performance test. Account access and managed organization settings can still restrict the steps below.
Check the client and connection first
Run claude --version. Anthropic documents v2.1.280 or later for Opus 5.5; its supported update command is claude update. On a managed installation, follow your organization's update process. See the official model configuration guide.
Next identify the connection: direct Anthropic access, a cloud provider, or a gateway. This matters because a provider can expose a different model catalog. Do not change credentials, endpoint, and model simultaneously: if the request fails, you would lose the ability to tell which change caused it.
Keep the previous working configuration available, without copying API secrets into notes or screenshots.
Select an explicit model, then verify it
For a supported direct Anthropic route, start a session with:
In an existing interactive session, use /model to choose the model, then /status to inspect the current selection. These are the selection and verification methods in Anthropic's setup instructions.
The shorter opus alias can resolve differently across providers and can change over time. Use your provider's supported identifier when connecting through a cloud deployment; an Anthropic model ID is not a universal deployment name. Also, opusplan is not a promise to use Opus throughout: it selects Opus for planning and Sonnet for execution. Alias behavior.
Record the selected model and any later switch notices. The model-routing verification guide explains why the requested ID alone cannot prove the identity of every subsequent call.
Check effort without assuming maximum is best
Effort controls reasoning depth. Opus 5.5 normally starts at medium in Claude Code, but explicit settings and organization rules can override it. Check the actual setting rather than relying on an old screenshot. Effort configuration.
At the API level, Opus 5.5 uses always-on adaptive thinking; switching thinking off is not a supported cost-control method. Model specification.
Our suggested first run uses the effective default. Increase effort only after identifying a failure that more investigation might address. A missing dependency, an incorrect test command, or an unavailable tool will not be repaired merely by selecting a deeper reasoning level.
If Opus 5.5 is missing or rejected
Use the symptom to choose the next check:
These are diagnostic categories, not proof that a particular account has an entitlement. Keep the error text, client version, provider, and requested ID together when asking for support; redact secrets.
For a purchase decision, use the separate Opus 5.5 vs Sonnet 5.5 cost comparison. Subscription allowances and provider API prices answer different questions.
Give the first run an acceptance card
Start from an independent repository copy with a known revision. Here is a task card to fill in, not a tested benchmark prompt:
First ask for the cause and proposed scope. Then authorize the bounded change. Review both the diff and the check output: a plausible explanation is not evidence that the bug is fixed. If the check cannot run, preserve that limitation instead of counting the task as a pass.
Save this card and result before moving to a larger task. They give a later Session something concrete to continue, rather than forcing it to reconstruct your expectations from a conversation.
Continue the project in Agent.Space
The commands above concern the Claude Code client. Agent.Space has its own compatible harness/model selector; this guide does not establish that Opus 5.5 is available for every account there.
Check the live selector and current pricing. If the supported combination fits, start an Agent.Space Workspace with the task card and project files. The Claude Code Workspace guide explains how to keep the task and review evidence with the project. Select the model deliberately, then judge the result against the task you actually needed done.