Model Reference

Model Reference

The model reference table: consult it when picking a model, update it when the fleet changes. The canonical copy lives here and nowhere else; the operator's always-injected rules file carries only the hard invariants and points here.

This is a lookup, not a step before every dispatch, and it never dispatches anything.

The table

Legacy illustrative scores, not a current fleet calibration. These entries lack dated evaluation provenance and are retained only to explain the axes. Do not route current work from these numbers or transfer fable-5's grade to Fable 5.1. A new model/effort needs its own operator-supplied calibration: date, task set, method, and observed trade-offs. Otherwise use available host capabilities and the session's current model under standing routing instructions.

Scores are 1–10, higher is better.

• Cost is subscription-limit burn, not dollars. The fleet runs on subscriptions, so a low cost score means "eats the weekly limit fast," not "expensive per token."
• Speed is wall-clock turnaround on the same task.
• Intelligence is how hard a problem the model can carry unsupervised.
• Code is coding craft: how correct and well-built the implementation comes out when the work is code.
• Taste covers user-facing surfaces only: UI/UX, copy, API shape, docs, research, audits.

One row per model, graded at the effort that model is actually run at. Effort is not a separate axis here, if you change the effort you habitually run a model at, re-grade its row rather than adding one.

Hard invariants

These carry the shape of each rule. The concrete policy, which models are in, which are out, where the floor sits: belongs to the operator's always-injected rules file, not to this skill; a plugin that hard-codes one operator's fleet ships a policy its adopters never chose.

• Orchestration stays home. Decomposing, dispatching, and judging a set of work always run on the session's own model, or on whatever the operator's rules file specifies, never on a weaker-model subagent.
• Standing escalation permission. When output misses the bar, rerun or redo on a smarter tier without asking. Judge the output, not the price tag: escalating costs less than shipping mediocre work.
• Cost and speed are tie-breakers only. When axes conflict for anything that ships, intelligence > taste > cost > speed. Neither of the last two buys a drop on the first two.
• Repo-local model policies override this table where they conflict.

Reading a currently calibrated table

The examples below apply only after the named models and grades have been verified for the current fleet; the legacy rows above do not establish that.

• Routine, well-specified work → the cheap end: luna only for truly mechanical bulk, sol for routine work that still needs judgment.
• Judgment-heavy, taste-critical, or silent-failure work → the frontier: fable-5, or opus-5 as the lower-burn lane. Putting high-tier judgment at the plan while a cheaper tier implements is often the better spend.
• A shipping taste surface needs taste ≥ 7: luna is not a taste route.
• Speed breaks ties, never quality. When two rows are level on the axis the work actually loads, take the faster one. It does not buy a drop on intelligence, taste, or code.

CI monitoring exception

Every CI watch uses a separate Opus (opus) agent on Claude or gpt-5.6-sol agent on Codex. Never use Astra or Fable to watch CI. Select the model explicitly; do not inherit a more capable parent. Parents already using Opus/Sol still delegate to a separate designated agent. Missing dispatch is a reported monitoring gap, not permission for parent polling or a prohibited fallback. Haiku and Sonnet remain prohibited. See fix-ci for the workflow.