Your Sonnet downshift for Checkout (acted Jun 28) is tracking 107% of projection.
778M tokens/mo realized vs 725M tokens projectedRoute 2,800 trivial Opus sessions to Sonnet across Checkout & DevEx
why · Rule ‘model over-provisioning’ fired: 2,800 sessions in 30d were ≤6 min, ≤10 tool calls, ≥85% acceptance — Opus adds no measurable quality on these.
better · Adopt a default-model policy: trivial edits start on Sonnet/Haiku, escalate only on failure. Ship the routing note and fold it into the onboarding playbook.
POLICYAdmin · one config changeSet policy →
Org cache hit 61% vs 75% target — worst offender SRE Core at 58%
14 stranded seats — activated, idle 30+ days
295M tokens of allowance leaked since posted 12d agoCheckout runs 1,525,000 tokens per accepted diff vs 960,000 tokens org median — driver: model mix
9 dangerous-command overrides on prod-adjacent repos this week — Sam Doyle, Leo Rao +2
Vendor Pod Chennai — Rahul M., Arjun K. +3: 120 Opus sessions on trivial tasks
Cleared this week: 9 moves → 2.0B tokens/mo projected (verification running on 4)
Compiled 07:30 from 1,204 sessions · quota weight table v3 (May 12) · measured vs proxy tagged · Accountability-focused