OpenAI Codex and ChatGPT Head: All Users Logged in with ChatGPT Accounts Can Now Use Auto-review for Free
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
ABAB AI Insight
Sottiaux's Codex initially compressed choices into two options: default sandbox requiring approval for everything, or Full Access removing supervision. In April 2026, OpenAI released Auto-review on the alignment page to fill this gap. He has since adjusted usage and reset multiple times through public posts, and the product rhythm itself relies on this account for external releases.
This free offering that does not deduct from the plan effectively shifts the cost of the second reviewing agent away from the user's quota. OpenAI is using its own computing power to extend agent sessions rather than selling an additional layer of security. The reviewing agent is separated from the executing agent, allowing the main agent to optimize task completion while the reviewing agent only monitors for boundary violations and intent deviations, avoiding the same model acting as both athlete and referee.
This approach is similar to Anthropic's use of permission pop-ups in Claude Code to gain enterprise trust, and it is also close to GitHub Copilot's confinement of agent actions within a repository sandbox. The difference is that OpenAI replaces the approver with another agent and directly offers the mode that has already accounted for the majority of internal token usage for free. The industry is transitioning from completion to long-task agents, and the ability to run tasks with minimal interruptions is becoming more crucial for retention than single-instance completions.
Essentially, this is a transfer of pricing power. Security reviews that once relied on user time for payment are now subsidized by platform computing power. With 200 times fewer manual interruptions, the ability to "run unattended" becomes a default capability. Once the quota is no longer consumed by reviews, users are more willing to keep long tasks in Codex rather than switching to non-sandbox tools to save on approval counts.
ABAB News · Cognitive Laws
If security relies on human approval, long tasks will stop.
Execution and review must be separated.
What is free is the quota; what is locked is the habit.