Claude Opus 5.5 makes premium coding agents cheaper, not simpler
Anthropic cut token and cache-read prices while preserving flagship coding performance, but the API changes make this a measured migration, not a drop-in upgrade.
The price moved
Anthropic released Claude Opus 5.5 on September 22 with a one-million-token context window. It costs $4 per million input tokens, $20 per million output tokens and $0.20 per million cache reads. Fresh input and output are 20% cheaper than Opus 5, while cache reads are 60% cheaper. It is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
Efficiency beats hype
The useful number is not price per token but work required per completed task. In SonarSource's Java evaluation, Opus 5.5 stayed within one percentage point of Opus 5's executable-test pass rate while using 40% fewer tokens and writing 27.5% less code. Bug density per line and concurrency findings increased, however, so shorter output should not be confused with automatic correctness.
The migration tax
This is not a safe model-ID-only replacement. Thinking cannot be disabled, forced tool selection can return an error, an older computer-use tool is rejected on some routes, and text between tool calls changes response shape. A production agent can therefore go quiet, alter its control flow or fail even when the request remains syntactically valid.
Our rollout
Virtual Arc would first run parallel evaluations on representative work and measure completed-task cost, retries, latency and human review time. We would route supervised multi-file changes, debugging and long cache-heavy sessions to Opus 5.5. Mechanical edits stay on cheaper models, and the existing route remains available until the new model proves its advantage inside our own harness.
Claude Opus 5.5 is the first recent premium-model launch we would treat as an economics event rather than a capability demo. The list-price cut alone is useful, but the operational shift is the 60% lower cache-read rate and independent evidence that it can complete comparable coding work with fewer tokens and less code to review. That combination can lower cost per completed task on long agent loops, where repeated context and output review dominate. Virtual Arc would start a production-shaped pilot now rather than wait for another model cycle: pin the model, update the harness for its new thinking and tool-call rules, and compare completed tasks, retries, latency and review time against the current route. We would not make it the universal default. Mechanical edits stay on cheaper models; Opus 5.5 gets supervised multi-file changes and investigations where fewer failed turns can justify the premium. If it wins on finished work rather than benchmark points, we route more traffic; if not, the migration stops at the pilot.