Claude Opus 5.5 lowers prices and changes the migration check

Anthropic released Claude Opus 5.5 on September 22 with lower token prices and claimed gains in coding and knowledge work. For an operational agent, the migration question is specific: does the new configuration preserve the exception that should stop an otherwise plausible action?

The release announcement lists input and output prices of $4 and $20 per million tokens, with cache reads at $0.20 per million. Anthropic says typical workloads at default settings cost 40% less than Opus 5. That is its workload claim, not a guaranteed reduction in a particular company's bill. The model is available through the Claude Platform as claude-opus-5-5 and through AWS, Google Cloud and Microsoft Azure.

Test the exception before the easy case

Evaluate the operational exception that makes a plausible answer unsafe before promoting the new model. Our earlier model-evaluation guidance makes that distinction: a convincing explanation and a usable operational decision can diverge.

In a hypothetical incident review, a recent exception forbids restarting a stateful service until replication catches up. The general runbook recommends a restart for the observed symptoms. Give both model configurations the same documents and require the answer to identify the conflict, preserve the hold and name the evidence needed to proceed. This is an illustrative test case, not a reported product failure.

Keep the evidence bundle, permitted tools and acceptance criteria fixed while comparing the model configurations. Count failed attempts and reviewer correction work alongside token charges. A cheaper completion is useful only when the workflow accepts it for the same reason it accepted the baseline.

Read the integration conditions

Anthropic also says Opus 5.5 uses preserved thinking for API accounts created on or after August 31, 2026, and cannot run with thinking switched off. Its announcement links the compatibility documentation. If an integration rewrites previous conversation context, test compatibility before routing live work to the new model.

Keep that compatibility check separate from the quality comparison. A successful request proves the integration can call the model; the fixed exception case tests whether its output remains acceptable. Retain the old routing configuration until both checks pass, and scope any production trial to the task and tool permissions actually reviewed. This publication has not independently benchmarked Opus 5.5.