Anthropic says serving test remapped Claude Code effort settings

• A “high” effort selection displaying as 10 triggered downgrade speculation after kimmonismus said the apparent value matched earlier “low” behavior and Fable felt weaker.
• Anthropic attributed the display to a serving-configuration test that remapped a numerical effort value, according to trq212's reply.
• The selected effort level still reached the model, and in-depth evaluations found no performance effect, trq212 said.
• The underlying models had not changed, trq212 wrote, while acknowledging plans to make regression reporting easier.

Anthropic’s public effort guide defines /effort through named levels, from low through max and auto. Its model-configuration docs also show that enterprise caps can substitute one named level for another, while the Claude Code release notes say Remote Control now synchronizes effort choices across web, terminal, Desktop, and VS Code sessions.

After kimmonismus reported that Claude Code showed “high” reasoning at 10/100, the post framed the number as evidence that the effective reasoning budget had been cut.

The report was about a status readout, not a benchmark or a reproduced task comparison. The number became meaningful to users because it appeared beside the named effort selection.

Anthropic said the number came from an API-serving experiment, not from a model change. In trq212's response, the company described the remap as follows:

• A serving configuration was being tested in Claude Code before rollout.
• That configuration mapped the numerical effort value differently.
• A session set to high could therefore report 10.
• The field was not a 0 to 100 scale.
• The displayed number had no standalone meaning.
• The selected named effort was still the effort being delivered.

That explanation confines the incident to the mapping between a user-facing label and an exposed internal numeric value. Anthropic did not say how the alternate mapping worked or how many users were assigned to it.

Anthropic made two separate claims: the models themselves were unchanged, and its in-depth evaluations found no performance regression. trq212 stated the first claim, while the earlier reply stated the second.

Neither response identifies the evaluation suite, sample size, task mix, or scores. The published Claude Code interface instead treats effort as a named control: Anthropic’s CLI reference lists low, medium, high, xhigh, and max, with availability depending on the model.

For users who believe they can reproduce a regression, trq212 asked for a /feedback report ID and said Anthropic would offer credits. The follow-up said the company was building a simpler way to report regressions in the moment in another reply.

The same account separately called Opus 5 “spiky” and said consistency and a warm Claude-like feel were a high priority. That is a broader product-quality acknowledgement than the numeric-effort remap, and it appeared alongside the feedback discussion.