The context · June 2026

Anthropic introduced Claude Sonnet 5 in June 2026, describing capabilities for work across coding tools, browsers and other software environments. [1]

A cheap action can start an expensive chain

An assistant may make several tool calls, retry a failed step and produce a result that still requires investigation. Looking only at the price of one model request misses this accumulated effort. The useful unit is a completed, accepted task.

Bound the attempt

Give a workflow a limit on time, tool use and repeated failures. Define the point at which it should return its findings to a person. A useful handover explains what succeeded, what failed and what evidence is available, so the reviewer does not have to start again.

Make stopping a normal outcome

Some requests lack enough information or require authority the assistant should not hold. Treat these as explicit workflow states rather than errors to conceal. A dependable assistant should be able to stop with a useful explanation. That behaviour matters as much as its performance on a successful demonstration.

Source & context

Anthropic · Claude Sonnet 5, 30 June 2026

This retrospective was written for the archive in September 2026. The linked primary source documents the announcement or event; the practical interpretation and proposed approach are Sansa’s editorial perspective. Public examples do not imply a client relationship. Product capabilities and guidance may have changed since the period discussed.

Another perspective · June 2026

A budgeted queue for routine research tasks

Continue reading

Working through a similar question?

Talk it through with Sansa