What Does “Recommended Set of Models” Mean in Codex?

What Does “Recommended Set of Models” Mean?

The “Default: Recommended set of models” selector means you're using OpenAI's recommended configuration instead of pinning a model yourself. OpenAI doesn't document it as a task router or one setting that works identically in every client. Here's what OpenAI says you can control in each:

• Desktop and Web Controls: In the ChatGPT desktop app and ChatGPT Work on the web, you start with your account's default Power setting. Move toward Faster or Smarter, or open Advanced to choose a specific model, reasoning effort, or speed.
• Local Configuration Files: The ChatGPT desktop app, Codex CLI, and IDE extension share the same config.toml configuration file. If you do not specify a model there, clients automatically use a recommended model.
• CLI Session Controls: In the Codex CLI, run /model to pick an available model and reasoning level, or pass model flags directly.
• Cloud Chat Behavior: For Codex cloud chats, OpenAI currently documents that you cannot change the default model.

What Model Does Codex Recommend by Default?

OpenAI now recommends GPT-6 Sol for complex coding and agentic workflows, GPT-6 Luna for focused tasks, and Astra for the hardest end-to-end work. Its illustrated starting preset is Sol Light. The actual default depends on your plan, client, workspace settings and rollout. OpenAI does not establish one universal default model ID for every account.

OpenAI recommends each model family for different work:

• Astra (gpt-6-astra): For the hardest end-to-end work. Choose Astra for complete workflows across code, apps, and research that need sustained reasoning and judgment.
• Sol (gpt-6-sol and gpt-6.1-sol): For complex, open-ended work. Choose Sol for ambiguous, difficult, or high-value tasks that need extra analysis, judgment, or polish, such as complex code changes, deep research, or polished documents. OpenAI introduced GPT-6.1 Sol on September 29, 2026 with near-Astra performance and set it as the bundled catalog default in Codex CLI 0.159.1.
• Terra (gpt-5.6-terra): The pragmatic all-rounder. Choose Terra for everyday work that needs strong reasoning and tool use when you do not need Sol's full depth.
• Luna (gpt-6-luna): For clear, repeatable tasks. Choose Luna for specific, high-volume tasks when you know what a good result looks like, such as extraction, classification, transformation, and structured summaries.

Note on deprecations: The gpt-5.4 and gpt-5.4-mini models retired from Codex with ChatGPT sign-in on August 31, 2026. Current OpenAI model guidance instructs developers to replace gpt-5.4 with gpt-6-sol and gpt-5.4-mini with gpt-6-luna in saved configurations, custom agents, and scheduled tasks.

OpenAI announced on September 14, 2026 that GPT-5.5 will retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026. If you use Codex with ChatGPT sign-in, replace gpt-5.5 with gpt-6-sol in workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks, and scripts when available to your plan and client before that date. Free and Go users are directed to GPT-6 Luna in the desktop app. This retirement does not apply to the OpenAI API.

Does Default Automatically Choose Between Astra, Sol, Terra, and Luna?

OpenAI does not currently document Default as dynamically choosing a different model for each task. Community discussions often speculate that “Default: Recommended set of models” checks prompt complexity and routes each task to a model, but OpenAI's documentation doesn't describe that behavior.

OpenAI documents starting with the default Power setting available to your account and adjusting the slider toward Faster or Smarter, while Advanced allows selecting a specific model, reasoning effort, or speed. Developers who want a specific model for a task should adjust the Power slider, open Advanced, or declare a model explicitly in configuration.

What Does the Power Slider Change?

Power trades reasoning depth against speed by stepping between preset model and reasoning-effort pairings. Start on whatever default your account has, then move toward Smarter for more reasoning or Faster for quicker, cheaper work.

OpenAI's current illustrations show Luna High, Sol Light (the starting preset), Sol Medium, Astra Light, Astra Medium and Astra Extra High. Some paid plans omit Astra Extra High. Open Advanced to choose a specific model, reasoning effort or speed; available options vary by account.

Does the Recommended Model Affect Codex Usage Limits?

Yes. Model and reasoning effort both change how fast you burn your 5-hour and weekly allowances. There is no published multiplier per preset, but heavier models and deeper reasoning go through allowance a lot quicker.

For example, OpenAI publishes different estimated 5-hour local-message ranges for each model. On ChatGPT Plus, its estimate for GPT-6 Astra is 5 to 45 local messages per 5 hours. Lighter models allow many more messages, within the cross-model 5 to 3,000 (Codex Usage summary of OpenAI's model-specific estimates) local message range. OpenAI still publishes Pro 5x and Pro 20x estimated-message columns. The Pro 20x column refers to the previous Pro $200 allowance, not every current subscription; the replacement allowance and Pro $500 multiplier are unpublished.

What Controlled Tests Showed About Model and Reasoning Choice

In an earlier two-task paired study using GPT-5.6 Sol and GPT-6 Astra, Sol and Astra each passed 57 / 57 deterministic checks. The displayed 5-hour meter moved by 3 points for Sol and 4 for Astra in aggregate. We reversed the model order between the two tasks.

In a later experiment, we ran the same frozen coding task on ChatGPT Plus once each on GPT-5.6 Luna, GPT-5.6 Sol, and GPT-6 Astra at Medium reasoning effort, graded by an identical deterministic 112-test suite:

• All three Medium runs — Luna, Sol, and Astra — passed all 112 tests.
• Sol finished fastest of the three (92.0s), ahead of Astra (109.7s) and Luna (118.2s).
• Measured 5-hour meter movement was lowest for Luna (no detectable movement at 1-point resolution), then Sol (2 points), then Astra (4 points).

A secondary run compared reasoning effort directly: GPT-5.6 Sol at Low reasoning also passed all 112 tests and finished about 16% faster than Sol at Medium (77.1s vs. 92.0s), and both moved the displayed 5-hour meter by 2 points. The 1-point difference in weekly meter movement between them is too coarse and too single-run to interpret. See the Low vs Medium measurement for the conditions.

How Do I Force Codex to Use a Specific Model?

To force Codex to use a specific model, choose it in Advanced, use the CLI's /model command or model flags, or save it in config.toml:

Frequently Asked Questions