Treat aliases as roles rather than permanent product names. If an alias is unavailable, select the nearest current equivalent and record the substitution.
| Level | Anthropic | GPT | Typical work |
|---|---|---|---|
| Host/orchestrator | Fable High | Sol High | Intent, decomposition, routing, evidence integration, final accountability |
| Host escalation | Fable Max | Sol xhigh | Genuinely difficult routing or integration after High is insufficient |
| Strong worker/reviewer | Opus High | Sol High | Complex implementation, debugging, design judgment, adversarial review |
| Planning/architecture worker | Opus High | Sol High | Bounded discovery, plan drafts, architecture options, risk analysis |
| Routine worker | Sonnet Medium/High | Sol Medium | Clear-spec coding, tests, migrations, refactors, tooling |
| Context scout | Sonnet or Haiku | Luna xhigh | Search, codebase mapping, dependency tracing, bounded research; escalate to Sol High when needed |
| Lightweight operator | Haiku | Luna low/medium | Chat, file organization, and tiny computer or clerical actions |
For Sol, move from High to xhigh before considering Max. Reserve Max for exceptional work and never select Ultra. Keep fast mode off unless the user explicitly values latency over quota.
- Balance Claude and Codex quota before preferring a provider; use task fit as the tie-breaker.
- Prefer Anthropic workers for frontend and visually sensitive implementation.
- Default orchestration to Fable High in Claude or Sol High in Codex. This does not make the host the default executor.
- Delegate bounded discovery, draft planning and architecture analysis, implementation, check execution, and review whenever practical.
- Escalate Sol orchestration from High to xhigh when routing or integration difficulty justifies it; reserve Max for exceptional cases.
- Work directly in the host only when the unit is truly atomic and dispatch overhead would exceed execution, required tools, authority, or context cannot be delegated safely, or one corrective redispatch has failed and escalation is exhausted. Record the exception.
- Split at file, module, service, or API boundaries; never split one coherent unit between workers.
- Parallelize only independent units with non-overlapping ownership.
- Default to depth one. Workers must not spawn subagents unless the user deliberately requests recursive delegation.
- Escalate after observed failure, not predicted difficulty.
- Stop after one corrective redispatch; on a second miss, escalate, then let the host take over only if escalation is exhausted.
- Check which provider has more usable quota.
- If quota is roughly balanced, prefer the provider whose strengths match the unit:
- Prefer Anthropic for frontend, visual polish, interaction judgment, and copy-sensitive work.
- Prefer GPT for backend, tests, migrations, broad refactors, tooling, and mechanical implementation.
- Choose the lowest tier that can reliably satisfy a frozen spec.
- Raise effort or tier only after an observed miss or when the risk class itself requires stronger reasoning.
Do not create artificial units merely to balance quota. Dispatch overhead is a valid direct-work exception only for a truly atomic unit, and the routing ledger must say so.