Operations
Compute Tiers — Opus is Tier A Only
Model-cost hygiene. Opus is reserved for manual strategic sessions and flagship foundation building. Automation runs on Sonnet; trivial pings run on Haiku. The split that keeps API compute from running away.
Compute Tiers — Opus is Tier A Only
THE RULE: Opus is never invoked autonomously or in a continuous loop. It is reserved for two things only — manual high-leverage strategic sessions, and flagship product foundation building. Everything else (background scripts, regeneration loops, admin tasks) routes through Sonnet. Trivial structural/handshake calls route through Haiku. Premium compute on a background loop is banned.
Why this exists (the 2026-06-24 cost-hygiene split)
Model IDs had drifted across the monorepo with no tier discipline — background regeneration scripts, a credential health-check ping, and the flagship install pipeline were all pinned to Opus indiscriminately. A max_tokens: 4 "hi" handshake was billing Opus rates. Background regeneration loops that need throughput, not deep reasoning, were on the most expensive model. There was no rule about which workloads earn premium compute, so cost tracked model-pin drift instead of leverage.
The tiers
| Tier | Model | What runs here |
|---|---|---|
| A — Opus | claude-opus-4-7 | Manual interactive sessions (strategy, positioning, deep architectural judgment) and flagship product generation: lib/install/foundation.ts, lib/install/components.ts, lib/install/generate-foundation.mjs, lib/install/qa-pass.ts. Brand voice and strategic output quality justify the cost here. |
| B — Sonnet | claude-sonnet-4-6 | Background administrative tasks and regeneration loops: scripts/regen-batch, scripts/regen-presell, scripts/regen-resources-page, scripts/blueprint-v3-regen, scripts/qa-pass. Non-interactive automation defaults here. |
| C — Haiku | claude-haiku-4-5 | Pure structural scaffolding, data emission, and handshake/health pings: lib/ops/credential-health.ts. Zero reasoning required. |
The standard
- Opus never runs autonomously. No cron, fan-out, queue, or background script may pin Opus. If it runs without a human in the loop, it is Tier B or C by definition.
- Flagship foundation is the one autonomous exception — and only because output quality is the product. The four
lib/install/*paths stay on Opus deliberately; downgrading them is a quality decision, not a cost decision, and is the CEO's call. - Automation defaults to Sonnet. Any new non-interactive script, regeneration loop, or admin task pins Sonnet unless a documented reason promotes or demotes it.
- Trivial calls default to Haiku. Health checks, structural scaffolding, and data-emission calls that need no reasoning pin Haiku. A
max_tokens-tiny handshake on a premium model is a defect. - No silent model drift. A new model pin states its tier. Pins are reviewed alongside the resource-discipline cost meters.
Acceptance
grepforclaude-opusin non-interactive scripts returns only the four flagshiplib/install/*paths — zero Opus pins inscripts/or health/ops utilities.- Every new automation pins Sonnet (or Haiku for trivial calls) by default; an Opus pin in automation requires an explicit documented exception.
- Model-cost is reviewed with Supabase egress + Vercel CPU as part of the monthly cost review.