Choose a model

Choose a provider first, then a model offered by its endpoint. The same model family can have different identifiers, availability, pricing, and transport requirements across providers.

Saved or session-only?

Action	When it applies	Saved for later?
`codai auth` model selection, then Save	Next session; first-run setup opens chat	Yes
`:preferences provider`, Default model, then Save	Next startup	Yes
Settings, Select model, then Enter	Future requests in the current mode	No
`:model agent <provider> <model>`	Future Agent requests in this session	No
`--model <id>`	This invocation	No

The session model selector applies immediately on Enter; it does not have a Save / Cancel confirmation. Escape closes it before selection. Provider setup and saved-default edits require Save.

Inspect or change the mode-specific selection:


:model
:model plan ollama qwen3-coder
:model agent openai gpt-5.4-mini

The compact form :model agent openai/gpt-5.4-mini also works. Changing the selection does not authorize a new credential or grant tool access.

Live discovery

Setup asks the selected endpoint for its model catalog where supported. It prefers that response to bundled suggestions, including for models released after Codai was installed.

Label	Meaning
Discovered model	Reported by the endpoint; not a guarantee of tool support
Saved model	Current configured selection retained in the picker
Suggested model	Bundled fallback; account access or installation is unverified
Custom model ID	A manually entered identifier

Successful discovery is not padded with bundled models. Failed or empty discovery uses suggestions where appropriate and still allows manual entry.

OpenAI-compatible APIs and Anthropic follow has_more / last_id pagination; Gemini follows nextPageToken. Ollama lists installed models. Discovery rejects repeated/missing cursors and catalogs beyond 100 pages or 10,000 entries instead of silently showing an incomplete list.

Duplicate and invalid IDs are removed, along with explicitly retired or non-chat entries where metadata supports filtering. Gemini entries must support generateContent; OpenRouter batch-only and non-text-output entries are excluded. Model listing is not a chat or coding-tool test.

Compatibility boundaries

Connection	Important limit
OpenAI API-key connection	Uses Chat Completions; direct GPT-6 tool calling requires Responses API
Anthropic native adapter	Multi-turn tool use requiring thinking-signature replay is not supported
Azure OpenAI	Enter a deployment ID, not a public model name; deployments are never guessed
Ollama	Select a model installed on the selected server
ChatGPT account	Uses account-specific models, separate from API-key fallback suggestions

Use the supported defaults when unsure. An explicitly saved model is not replaced when a bundled default changes.

Bundled fallback families

Fallbacks are convenience suggestions, not an exhaustive catalog. Account, region, endpoint, and model lifecycle can change availability.

Provider	Additional suggestions
OpenAI	GPT-6 variants, GPT-5.5 / 5.4 / 5.2 / 5.1 / 5, GPT-4.1 variants
Anthropic	Sonnet / Opus / Haiku 5.5, Fable 5.1, 5-series and selected 4-series models
Gemini	4 Argon, 3.6 Flash, 3.5 / 3.1 Flash Lite, Pro / Flash previews, 2.5 legacy models
DeepSeek	`deepseek-v4-pro`
OpenRouter	Provider-qualified OpenAI, Claude, Gemini, DeepSeek, Qwen, Mistral, and Grok families
Qwen	Qwen3.8 Max / Flash / Omni Flash, Qwen3 Coder, Plus / Max / Flash aliases
Mistral	Large 4, Small 4, Codestral and Large / Small aliases, Ministral variants
Grok	4.6 / 4.5 / 4.3, 4.20 variants, Grok Build 0.1

Azure has no guessed deployment suggestions. SDK hosts can inspect ProviderDescriptor.DefaultModel and SuggestedModels for exact bundled IDs.

Next steps

Usage and pricing explains estimates and unknown model prices.
Troubleshooting covers discovery and transport failures.