Thibault Sottiaux, head of OpenAI Codex and ChatGPT, announced that all users logged in with ChatGPT accounts can now use Auto-review for free. The switch is located in settings, permissions, Auto-review, and the feature does not consume plan usage.

This capability modifies the default sandbox. In default mode, almost every step that goes beyond limits requires human approval, which can lead to decision fatigue in long tasks unless a large number of rules are written in advance. Auto-review allows the main agent to continue working while a second agent reviews all actions taken.

The reviewing agent has one goal: to block high-risk actions and operations that deviate from the user's original intent. It is not responsible for completing the task. Official documentation states that this is a replacement for the reviewer, not an opening of permissions. The main agent remains within the original sandbox, and the boundaries of the network and file system remain unchanged.

The requests that will be reviewed are those that go beyond limits: commands requesting higher sandbox permissions, network access blocked by policies, file modifications outside writable directories, and tool calls that require approval as per annotations. If approved, execution continues; if denied, the main agent is required to take a safer path or stop to ask the user. In full release mode, there are no items to review.

Internal data previously disclosed by OpenAI indicates that in Auto-review mode, the number of times a session stops waiting for human approval is about 200 times less than manual approval. Among the few actions that require review, the reviewing agent approves about 99%. Most of the token usage from the Codex desktop version within the company has come from this mode, and the proportion is still increasing.

The buyer is Codex users logged into ChatGPT, and the seller is OpenAI, which has removed the review computing power from the plan quota. This is a product day update, not a new model release. Usage is no longer deducted from the plan, friction costs for long tasks decrease, benefiting heavy users and teams running unattended tasks. The default process, which still requires manual approval for each step, is under pressure; if competitors continue to make security a repeated pop-up, it will also be compared.

Sottiaux marked this as Day 2.1. Review only takes effect under interactive approval policies. Users can still revert to approving each item individually or switch to full access without further inquiries.

Source: Public Information