Skip to content

Build Notes

GPT-6.2 in Codex: Choosing Astra or Sol for Your Next Task

By Miles Park·

GPT-6.2 Astra and GPT-6.2 Sol arrived in ChatGPT Work and Codex on November 17, 2026. Astra is aimed at demanding projects; Sol is the everyday choice for coding, writing and multistep work. The useful first step is choosing a model for a specific task and defining how you will review its result.

Which model should you start with?

Your taskStarting pointWhat to review
Trace a failure through several services, implement a fix and test itGPT-6.2 AstraEvidence for the cause, scope of changes and test results
Edit a small function, draft documentation or organize project notesGPT-6.2 SolCorrectness, clarity and adherence to your constraints

These starting points follow OpenAI’s positioning, not a benchmark we ran. The announcement also says Sol can complete more tasks on the same paid plan than GPT-6.1 Sol; it does not give a universal number of extra tasks.

Select GPT-6.2 in the app or CLI

You need access through a supported plan and client. In the desktop app, open the model and reasoning control beneath the composer. If you see the Power slider, open Advanced to select a specific model. Choose an available GPT-6.2 model and start with the selected reasoning effort.

In an interactive Codex CLI session, enter /model to choose a model or adjust reasoning effort. OpenAI documents this command for starting a new Sol session:

codex --model gpt-6.2-sol

Model availability and reasoning effort are separate from speed. If a model is absent, check the official availability documentation and your workspace’s access settings before changing configuration.

Two prompts with a clear finish line

For a focused Sol task, try this in a repository you are permitted to work on:

Review the CSV email-import validator and its tests.
Suggest a small improvement for duplicate email handling.
Explain the edge cases and proposed test cases before editing files.

For an Astra investigation, describe the symptom and ask for evidence before implementation:

Investigate why a saved email import can appear twice after a retry.
Trace the request, persistence and retry paths. Identify the cause with
file references, propose the smallest fix, then run relevant tests.
Report which checks passed and what remains unverified.

These are our illustrative prompts, not executed demonstrations. Review the proposed diff and actual test output. A plausible explanation alone does not establish the cause. OpenAI’s prompting guide similarly recommends specifying the criteria you will use to judge the completed work.

When to consider a faster speed

Fast and Ultrafast generate tokens faster on the same model and consume more allowance or credits. In the app, the lightning-bolt control opens a speed menu when Ultrafast is available; in the advanced picker, use Speed. The CLI documents /fast and /ultrafast; entering the active command again returns to Standard.

Ultrafast availability depends on your plan and, for managed workspaces, model access and a separate permission. Check the current requirements before selecting it. Faster generation does not verify code quality or replace tests.

Sources and related guides

Official references: November 17 announcement, release announcement, model selection, Fast and Ultrafast, and GPT-6.2 prompting guidance.

For terminal navigation, see our side-chat and task-search walkthrough. Browse original announcement dates in Codex Update history.

Join the discussion

Your email address will not be published. Required fields are marked *