Per-model thinking level (a default reasoning level per model)

What it is

One-paragraph answer: "Per-model thinking level" is a setting (Settings → Session, just under the single Thinking Level) that lets you give each model its own default reasoning/effort level — e.g. Opus → Max, Haiku → Low, Codex GPT-5.5 → High — so every new session inherits the right level based on which model it uses, without you setting it by hand each time. It's a supplement, not a replacement: any model left on "Inherit default" falls back to your single Thinking Level (for Claude) or the model's own default (for Codex), so nothing changes until you customize a model. It covers every engine that has a reasoning knob — today Claude and Codex — and your per-session and per-project choices always win over it.

Where to find it

Where to find it

• Settings → Session → "Per-model thinking level" (the card directly below Thinking Level).
• It lists one row per current model of each engine that exposes a reasoning knob: Claude (Fable 5.1 / Fable 5 / Opus 5.5 / Opus 5 / Sonnet 5.5 / Sonnet 5 / Haiku 4.5) and Codex (GPT-6.1 Sol / GPT-6 Astra / GPT-6 Sol / GPT-6 Luna / GPT-5.6 Sol / GPT-5.6 Terra / GPT-5.6 Luna / GPT-5.5). Legacy Claude versions aren't listed here (you can still override them per-session).
• Each row is a dropdown: "Inherit default (…)" plus that engine's own levels — Claude offers Auto / Low / Medium / High / Max; Codex offers Minimal / Low / Medium / High.
• Searchable in the Settings search bar as "per-model thinking level" (also "per model", "reasoning", the model names).
• Or set one model at a time from a fresh session — pick a non-default thinking level, then open the Thinking pill's dropdown: below the level list sits a small defaults editor with one row per scope — This model, All (engine) sessions / All (harness) sessions (named for the actual engine / harness), Only this project, and Everywhere (Claude's flat level). Each row shows what it's currently set to, tags the one that's actually in effect (and marks any outranked row "Overridden above"), and gives you Set / Change / Clear — so This model writes that model's per-model default (this same map), the quick way to set the level for the model you're looking at, and you can save at whatever breadth you want. See start-a-new-session.md.

How it behaves

Why it exists

Different models have different sweet spots. Opus benefits from maximum reasoning; Haiku is meant to be fast, so heavy thinking is wasted on it; Codex has its own effort scale. Before this, Omniscio had one thinking level for all Claude sessions, so you had to override thinking by hand whenever you switched models. Now you set it once per model and forget it.

"Inherit default" vs "Auto" — they're different

• Inherit default removes the model's own setting, so it falls back down the chain (to your single Thinking Level for Claude, or the model's own default for Codex). Pick this to un-customize a model.
• Auto (Claude only) is an explicit choice that pins the model to the model's own default reasoning — it ignores your single global Thinking Level. Pick this when you specifically want "let the model decide" for that model even though your global default is, say, Max.

What wins (precedence)

your per-session pick → the project's default → the per-model default → the per-provider default → the per-harness default → your single global Thinking Level

The per-model default only fills in when you haven't chosen something more specific for that session or project. A per-project thinking default (set in Edit Project, or via the fresh-session Thinking defaults editor's Only this project row) still wins over all of the per-model / per-provider / per-harness defaults, because a project setting is the narrower, more deliberate scope. The per-provider and per-harness layers (set from that same defaults editor) sit just under the per-model default and above your single global Thinking Level.

Both engines, and the safety net

• Claude sessions read the level through CLAUDE_CODE_EFFORT_LEVEL. The per-model default folds into the "global" slot alongside the newer per-provider and per-harness defaults — the order inside that slot is per-model → per-provider → per-harness → your single Thinking Level — and the whole fold still loses to a session/project pick.
• Codex sessions read it through their model_reasoning_effort config. Codex has no single global thinking level of its own, so the per-model default is simply its lowest layer (session → project → per-model).
• Validation: each engine only offers levels it actually supports, and the resolver re-validates at read time — so a Claude-only level like Max can never leak onto a Codex model, and a Codex-only level like Minimal can never leak onto a Claude model. An invalid or stale entry is silently dropped, never sent to a session.

What is never affected

Engines with no reasoning knob (Gemini, DeepSeek, Kimi, GLM, MiniMax, Meta, Grok, Cursor, OpenCode, Pi, Anti-Gravity, Hermes) don't appear here and are unaffected — Cursor bakes effort into the model id, and the rest expose no effort control. Existing installs are unchanged until you set a per-model level (the setting starts empty), so there's no migration and no surprise.

For agents

For agents with repo access — where this lives

• Setting: modelDefaultThinkingLevels (a modelId → level map, default {}) — plus the sibling conditional-thinking maps providerDefaultThinkingLevels (providerId → level) and harnessDefaultThinkingLevels (toolId → level), both default {} — in AppSettings, schema in src/shared/types/settings/accounts-providers-settings.ts.
• Resolution: src/shared/providers/provider-default-model.ts — getModelDefaultEffort / getProviderDefaultThinking / getHarnessDefaultThinking (the validated per-scope lookups) and resolveModelGlobalThinking (the one shared fold — per-model → per-provider → per-harness → flat-global — that the spawn and the chip both read, so the pill can't drift from what launches; toolIdForProvider in tool-grouping.ts keys the harness layer). Effort-capable engines come from providersWithReasoningEffort in src/shared/providers/provider-models.ts.
• Claude spawn: resolveSpawnThinking in src/main/process/spawn-build.ts (takes the resolved model, folds the per-model default into the global slot; the model param is required so every spawn path — local / SSH / aside — supplies it). Codex spawn: ensureStarted in src/main/services/engines/codex-session-manager.ts.
• UI: the PerModelThinkingSection card in src/renderer/src/features/settings/sections/session/SessionSettings.tsx writes the map, as does the fresh-session "Save as default" banner (SetDefaultHint, rendered below the Model/Thinking pickers) — This model via buildModelDefaultThinkingPatch, This provider / This harness via buildProviderDefaultThinkingPatch / buildHarnessDefaultThinkingPatch, This project / Everywhere via the project + flat writers — wired in src/renderer/src/features/sessions/SessionModelThinkingChips.tsx; the chip also shows a Project / Session source badge, and reads the same resolver there; search entry in src/renderer/src/features/settings/SessionsSettings-search.ts.
• Contract: .claude/memory/contracts/provider-registry-contract.md (PART 3, invariant the-per-model-effort-default-is-validated) + start-config-staging-contract.md (set-default-hint-is-the-one-surface — SetDefaultHint is the one surface for a Model/Thinking default; the per-level editor superseded-per-level-editor introduced is retired).
• Tests: tests/unit/shared/provider-default-model.test.ts, tests/unit/process/spawn-build.test.ts, tests/unit/codex-session-manager.test.ts, tests/unit/features/settings/per-model-thinking-section.test.tsx.

Related

• model-vendors.md — who pays for a non-Claude session and which company serves it, the other choice Settings makes for you without changing the model.
• ai-providers.md — the models and engines this setting covers.
• per-window-zoom.md — another per-scope default that stacks with a global one.