Escalate to Opus by need, not by agent: P0s, planning, and any pull sent back twice run on Opus; everything else on the agent's own model #7570

Measured over 24 h, 2026-10-27:

• 466 forge pulls landed, for about $2,016 at list price (about $4.30 per landed pull).
• Opus: 1,399 runs, $1,642 ($1.17 a run). Sonnet: 1,517 runs, $297 ($0.20 a run).
• Failed or cancelled runs cost $18.
• Reviews: 601 approved, 297 changes requested.
• 151 of the 466 landed pulls are "A test pins…".

The owner asked to act on the findings. Builders and the sentinel reviewer were moved to Sonnet at 22:50Z; torque and ledger stay on Opus.

Build:

• A run picks its model from the work, not only the agent: P0 or label opus → opus, planner runs → opus, a pull with 2+ changes-requested rounds → opus for its next round; otherwise the agent's model.
• The run row records the model and why (model_reason).
• The scorecard shows spend and landed pulls per model per day, so the switch is measured.

Tests: a P0 issue run picks opus; a twice-bounced pull's next run picks opus; an ordinary run picks the agent's model.

Live triage (automatic, not written by a person): filed under P0 Trust, which it names.