OpenAI Publishes GPT-6.2 Guide for Codex: When to Use Medium, High, Extra High and Ultra
A Codex agent can waste time on a trivial fix or give a shallow analysis of a hard bug, and the reasoning level you pick often decides which. OpenAI has published a guide to GPT-6.2 (Astra and Sol): Medium for a quick overview, High for a bug that spans several systems, Extra High when several factors interact at once. In Ultra mode, Codex decides on its own when to call subagents. Assign the task as you would to a colleague, with an acceptance criterion.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
A coding agent in Codex can spend hours fiddling with a simple edit, or handle a hard bug at the boundary between services only superficially. Often the model itself is not the cause. Two things matter more: which reasoning level you enabled and how you framed the task.
OpenAI has published a practical guide on learn.chatgpt.com for working with the GPT-6.
2 family (the Astra and Sol models) inside Codex. It describes three reasoning-depth settings — Medium, High and Extra High — plus an Ultra mode, in which Codex itself decides when to launch subagents. That makes four options in total, and below we cover when to pick each.
This is a retelling based on the Telegram channel "Метаверсище и ИИще"; the guide itself is linked at the end.
Assign the task like you would to a colleague, and give acceptance criteria up front
The main recommendation: phrase the task the way you would for a human colleague on the project, and always state acceptance criteria, meaning a condition by which the result can be checked. The guide's example: an icon must stay legible in a toolbar at 24 pixels. The agent can verify a criterion like that itself, and you don't have to keep sending it back with "still not right."
For developers, this carries over to any kind of work. "Fix the test" is a weak assignment. "Fix the test without changing the public API, and run only the affected package" works: the agent has a boundary and a clear point at which it can stop.
Which reasoning level to choose and when to call subagents
The guide advises matching the reasoning_effort parameter, which controls how much the model "thinks" before answering, to the complexity of the investigation:
- Medium — a quick overview of code or a problem;
- High — when you need to trace a cause across several systems;
- Extra High — when several factors interact at once and there is no simple explanation;
- Ultra — Codex itself decides when to launch subagents.
In practice, Medium is enough for routine work, and it is worth raising the level where a bug is intermittent and touches several services.
Subagents are separate instances of the agent that take on independent parts of the work in parallel. For example, one is told to update the documentation, another to check a module, and a third to analyze logs. If a task has independent branches like these, the guide advises adding the directive "use subagents as needed" explicitly to the prompt, a skill or the AGENTS.md file. In Ultra mode, Codex decides on its own whether to launch subagents.
Visual tasks and AGENTS.md hygiene
For slides, SVG and 3D scenes, the recommendation has three steps. First, give the model a reference (or generate one with GPT Image). Then make it render the result. Finally, check the image at its final viewing scale. An icon that looks good enlarged can turn into a smudge at 24 pixels. The SVG & 3D Design plugin is available for editable illustrations and 3D models.
AGENTS.md is a file with standing instructions for the agent in a repository. The recommendations for it are:
- keep the file compact;
- move personal preferences to ~/.codex/AGENTS.md;
- when several agents work locally, limit the number of simultaneous builds and tests;
- delete temporary files and background processes when done.
The last point will be familiar to anyone who has watched several agents start builds at once and their laptop slow to a crawl.
What to keep in mind
This is a short retelling, not the full text of the guide, so it is better to check the details against the original. The channel's author also adds a personal observation: the new Astra can be stubborn. It solves the task the way it understood it, and reacts rather calmly to a user's indignant shouting. This is one person's impression, not a measurement. But it shows why the guide insists so much on acceptance criteria: the more precisely the goal is described at the start, the less you have to argue with the model along the way.
The practical takeaway is simple. Before launching, choose a reasoning level to fit the task, write down verifiable conditions for completion, explicitly allow subagents where the branches are independent, and keep AGENTS.md short.
- Nov 20, 2026Model ML Raises $250 Million for Banker-Focused AI Agents at a $1.5 Billion Valuation
- Nov 20, 2026Yandex 360 Opens 7 Office Services to AI Agents via MCP Server
- Nov 20, 2026Tencent open-sources an agent that detects generated video with 95.2% accuracy
- Nov 20, 2026US Officials Discuss Forcing Companies to Disclose Chinese AI: Chinese Models Take All Top 3 Spots on OpenRouter
Key Questions Answered
What are the key details on Assign the task like you would to a colleague, and give acceptance criteria up front?
The main recommendation: phrase the task the way you would for a human colleague on the project, and always state acceptance criteria, meaning a condition by which the result can be checked. The guide's example: an icon must stay legible in a toolbar at 24 pixels. The agent can verify a criterion like that itself, and you don't have to keep sending it back with "still not right." For developers, this carries over to any kind of work. "Fix the test" is a weak assignment.
Which reasoning level to choose and when to call subagents?
The guide advises matching the reasoning_effort parameter, which controls how much the model "thinks" before answering, to the complexity of the investigation: Medium — a quick overview of code or a problem; High — when you need to trace a cause across several systems; Extra High — when several factors interact at once and there is no simple explanation; Ultra — Codex itself decides when to launch subagents.
What are the key details on Visual tasks and AGENTS.md hygiene?
For slides, SVG and 3D scenes, the recommendation has three steps. First, give the model a reference (or generate one with GPT Image). Then make it render the result. Finally, check the image at its final viewing scale. An icon that looks good enlarged can turn into a smudge at 24 pixels. The SVG & 3D Design plugin is available for editable illustrations and 3D models. AGENTS.md is a file with standing instructions for the agent in a repository.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.