Reasoning effort

Separate from which model answers is how hard that model thinks. A tool-resumption turn, where the model asked for a file and is now being handed it, rarely needs extended reasoning. A new question, or a turn that follows an error, usually does.

tokunseba tier enable 3
tokunseba config set tier3.effort_routing true

What it changesCopy Anchor Link

The turn is otherwise untouched. No text is rewritten, no model is swapped.

When it firesCopy Anchor Link

Only when all of these hold:

• tier 3 is on and tier3.effort_routing is on
• the judge scored the turn as easy, above the confidence gate
• the turn is not sensitive
• the request did not already specify an effort setting of its own

If the client asked for a specific thinking budget, that is a decision the client made and tokunseba leaves it alone.

Seeing itCopy Anchor Link

effort_set   23   an easy turn was sent at low effort

tokunseba stats --since 7d

Because this is tier 3, sessions are split into control and treatment arms, so:

tokunseba stats --ab

will tell you whether the treatment arm actually finished tasks in fewer requests or more. That is the number that matters: a turn that thinks less and then needs two more turns has not helped.