Model catalog

Checked against the provider documentation on 2026-09-24. These are API model IDs and documented effort values. Availability depends on your account, installed native CLI, and authentication path. Exclude models your plan cannot use. Nothing here establishes cache preservation.

Tool	Default tier	Family	Provider model ID	Documented API effort values
Claude	Routine	Haiku	claude-haiku-4-5-20251001	No effort parameter
Claude	Standard	Sonnet	claude-sonnet-5	low, medium, high, xhigh, max
Claude	Complex	Opus 5.5	claude-opus-5-5	low, medium, high, xhigh, max
Claude	Demanding	Fable	claude-fable-5-1	low, medium, high, xhigh, max
Codex	Routine	GPT-6 Luna	gpt-6-luna	none, low, medium, high, xhigh, max
Codex	Standard and complex	GPT-6 Sol	gpt-6-sol	none, low, medium, high, xhigh, max
Codex	Demanding	GPT-6 Astra	gpt-6-astra	low, medium, high, xhigh, max

Previous-generation models remain in the catalog so personal policies can select them. They are not routed by default:

Tool	Family	Provider model ID	Documented API effort values
Claude	Opus 5	claude-opus-5	low, medium, high, xhigh, max
Codex	GPT-5.6 Luna	gpt-5.6-luna	none, low, medium, high, xhigh, max
Codex	GPT-5.6 Terra	gpt-5.6-terra	none, low, medium, high, xhigh, max
Codex	GPT-5.6 Sol	gpt-5.6-sol	none, low, medium, high, xhigh, max

GPT-6 has no Terra model, so GPT-6 Sol serves both middle Codex tiers through two profiles: codex-sol-balanced defaults to medium reasoning when effort is uncertain, and codex-sol defaults to high. The built-in codex-terra profile keeps GPT-5.6 Terra selectable while Codex offers it.

All four tiers are active in the shipped policy. Automatic effort selection uses low, medium, high, xhigh, and max for the effort-capable models; Haiku gets no effort parameter. The catalog validates allowed values. It does not establish account access or cache-preserving effort changes. Invalid/unavailable classification or insufficient context uses the configured uncertain profile, which defaults to Opus 5.5/Sol at high. Low model confidence instead imposes a balanced minimum while retaining stronger proposals; see routing policy.

The Claude IDs and Haiku’s lack of an effort parameter are from the model overview. Effort support is from the effort reference. Effort interacts with thinking configuration and output limits: an allowed effort value alone does not prove the whole native request is valid. In particular, Fable and Opus 5.5 always think: neither accepts disabled thinking. The Claude adapter sends adaptive thinking with every effort value, and sends no thinking or effort fields to Haiku.

OpenAI’s documented effort values come from the individual GPT-6 Luna, GPT-6 Sol, and GPT-6 Astra model pages, and the earlier Luna, Terra, and Sol pages. A setting exposed by a particular Codex app build is not automatically a supported value on every API/authentication path. For example, Codex 0.156.1 lists low through max for these models (plus ultra for Sol and Astra) and does not list none; Switchboard only sends low through max. GPT-6 Luna and Sol require a Codex build that bundles them; switchboard doctor checks your installed Codex.

The bundled compatibility list is intentionally empty. A verified adaptive-effort entry needs evidence for the exact model, tool version, authentication path, mode, update mechanism, and continuation/resume behavior. Until such support exists, both model and effort stay fixed for the conversation. See conversation and cache behavior.