What Claude Sonnet 4.5 introduced

01Separate the model from the tools around it

A model's reasoning capability, a coding interface's rollback controls and an application's memory system are different parts of a workflow. The announcement discussed all of them, so a claim about the overall experience should identify which part changed. ^[1]

Anthropic's benchmark results were reported evaluations under particular conditions. They did not guarantee that every repository or long-running task would improve.

When comparing models, use the same task and acceptance criteria. Review the patch, run the project's relevant checks and count the corrections required. For work involving external systems, define the permissions and stopping conditions before allowing an agent to continue unattended.