2026-10-13-codex-default-model-moved-again-in-a-hotfix-and-fresh-tui-threads-now-take-the-server-s-defaults

Default model moved again in a hotfix, and fresh TUI threads now take the server’s defaults. For the second window running, a patch release changed what runs when model is unset: 0.159.1, published the same day as the changelog’s “GPT-6.1 Sol in Codex and ChatGPT Work” entry, made gpt-6.1-sol priority 1 in the bundled and Bedrock catalogs, and 0.161.0 carried the same default on the main line. Its catalog entry defaults to low reasoning and lists ultra as its top effort, described as “Maximum reasoning with automatic task delegation” (the effort already existed for other models at 0.155.1). 0.156.1 had already put GPT-6 Sol and GPT-6 Luna in the picker and made the rate-limit prompt recommend Luna. In 0.162.0 fresh TUI threads omit unselected model, model_provider and model_reasoning_effort from thread/start, so the connected server’s configuration decides: the local daemon’s config/read, managed new-thread defaults, or its catalog when no model is configured (#50013, #50913). The picker stopped appending “(default)” to any entry (#51235). The legacy thread and turn multiAgentMode fields are deprecated and ignored in favor of ultra effort; the server discards the value, and this already held at 0.155.1, so it is a correction to the prior profile rather than a new in-window removal. The catalog is “backed by bundled models, cache, and /models” (https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/models-manager/src/manager.rs#L285), so an unset model can still move without an upgrade. GPT-5.5 is still listed at 0.162.1 with an upgrade prompt to gpt-6-sol and retirement_at: null (models.json L1397-L1406), while the changelog entry of 2026-09-14 retires it “from ChatGPT, ChatGPT Work, and Codex on all plans” on 2026-10-14 and says the retirement “does not apply to the OpenAI API”. Managed requirements gained model_provider and model_providers, which override local and session selection (#44650, 0.156.0); the managed-configuration docs page (retrieved 2026-10-13) does not list them.

Channel: tagged-release; docs-only (changelog entries of 2026-09-14, 2026-09-22 and 2026-09-29). Half: both. Date: 2026-09-23 (0.156.1 GPT-6 Sol and Luna); 2026-09-29 (0.159.1 default GPT-6.1 Sol); 2026-10-08 (0.162.0 server defaults, no default label).

Operator consequence: Set model and model_reasoning_effort explicitly, in config.toml or on the command line, for anything with a cost, latency or eval baseline. A value in config.toml is part of the server’s configuration and survives the new precedence; an unset one follows the catalog. Re-run evals on fleets that floated through 0.159.1. Remove gpt-5.5 from configs, custom agents and scheduled tasks before 2026-10-14; the bundled 0.162.1 GPT-5.5 entry has no retirement timestamp; server or cached catalog metadata can differ. Admins who need a fixed gateway can now require it with model_provider in requirements.toml. The picker no longer tells anyone which model is the default, so the status line is where to look.

Receipt

• https://github.com/openai/codex/releases/tag/rust-v0.159.1
• https://github.com/openai/codex/blob/rust-v0.159.0/codex-rs/models-manager/models.json#L4-L74
• https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/models-manager/models.json#L177-L248
• https://github.com/openai/codex/pull/50013
• https://github.com/openai/codex/pull/51235
• https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/models-manager/models.json#L1397-L1406
• https://learn.chatgpt.com/docs/changelog
• https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/app-server-protocol/src/protocol/v2/thread.rs#L110-L113
• https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/app-server-protocol/src/protocol/v2/turn.rs#L284-L287
• https://github.com/openai/codex/blob/rust-v0.162.1/codex-rs/app-server/src/request_processors/thread_processor.rs#L1156