Codex Head Tibo: Automatic Approval No Longer Counts Towards Usage
Tibo, head of OpenAI Codex, summarized the second day updates, incorporating "Approve for me" (automatic approval) into the package without consuming usage, while also launching a simplified developer interface, meeting notes integration, and the Decisions API open to developers.
The automatic approval feature is available to users logged in with a ChatGPT account. The main agent still operates within the original sandbox, while out-of-bounds commands, blocked network requests, changes outside allowed directories, and applications or MCP calls requiring approval will be judged by a second review agent based on the user's original intent. High-risk or misaligned actions will be rejected and asked to take a safer path or stop to consult a person. The entry is selected from the permissions menu below the input box under "Approve for me." Tibo stated that once enabled, it would account for about 2-10% of the package, but it will no longer deduct usage. This feature was set as default internally at OpenAI in May, with approval counts reportedly dropping to about 1/200 of the original.
On the developer side, the five paid usage tiers have been consolidated into three: Build, Launch, Grow. The new highest tier, Grow, requires a cumulative interface payment of $500 to qualify, down from the previous threshold of $1000. Tibo specifically mentioned that this aims to reduce friction during interface development.
Meeting notes have been integrated into the ChatGPT desktop client. The Meetings plugin is in public beta for Pro and Business users on macOS, with Enterprise access to follow. The plugin takes notes and generates summaries and next steps based on user interactions and existing work, storing them in ChatGPT Space, which can be private or shared with teams, and then handed over to ChatGPT for project planning or drafting follow-ups. Tibo's usage involves directly embedding colleague conversations into context, then loading them into Codex for execution.
The Decisions API is now open to developers and is in public testing, expected to transition to the official version in a few weeks. Currently, it only connects to gpt-6-luna, using the dedicated POST /v1/decisions. Inputs can be text or images, with questions limited to true/false judgments, fixed options, or scoring scales, returning named answers. Official documentation states it is about 10 times faster than the Responses API. Tibo previously mentioned that end-to-end processing could be compressed to within a few hundred milliseconds and indicated that this real-time general classifier would be used back in their applications to improve the experience.
This is a usage benefit within the product, not external funding inflow or outflow. Review calls that previously consumed package quotas are now covered by OpenAI itself, offsetting the reasoning costs to allow long tasks to continue without waiting for approvals. Beneficiaries include ChatGPT logged-in users and developers aiming for higher interface tiers. The pressure is on the general conversational interface that still charges per use, as well as on proxy tools that rely on manual approvals and slow classification.
Source: Public Information
ABAB AI Insight
Tibo has made automatic approval the default internally at OpenAI since May, based on the alignment of the team’s review agents, rather than directly opening sandbox permissions. The same team previously released the Decisions API with Luna as a constraint for decision-making in September, and clarified at the end of September that the main Codex dot is included in the package without deducting usage. The path is quite fixed: first, reduce the number of approvals internally, then change the billing items to included items, and finally make the classifier an external interface.
The money has not flowed into acquisitions but has shifted the cost of review reasoning from user packages to OpenAI's own reasoning budget. The interface tiers have been consolidated from five to three, and the Grow threshold has been lowered from $1000 to $500, which reduces the entry barrier for accessing higher-speed tiers, encouraging developers to route classification, routing, and next steps to their own endpoints. Meeting notes go into Space and are then loaded into Codex, keeping the context from the desktop in their own workspace rather than handing it over to external note-taking tools.
This aligns with Anthropic's approach of making confirmation steps configurable permissions and GitHub Copilot's practice of keeping proxy actions within the repository sandbox: agents entering the chat box will modify files and execute requests. The industry has moved past the demonstration phase and is now in the control phase. Whoever removes manual approval from the main path first will be better positioned to handle long tasks. The difference is that OpenAI is now bearing the costs of the review model.
Structurally, this represents a shift in pricing power. General generation is sold per token, while real-time true/false judgments, options, and scoring have been separated into faster dedicated channels, with the document indicating a speed difference of about 10 times. The approval layer has also been removed from user quotas. The mechanism is that slow and expensive generation is no longer suitable for gatekeeping; gating and routing are now handled by smaller, faster classifiers, with the platform using free reviews to lock in default execution, and then lowering the tier thresholds to keep developers' decision traffic on their own interfaces.
ABAB News · Cognitive Laws
Free is for gatekeeping, billing is for generation.
Reducing approval counts allows for longer task durations.
The platform benefits from review, trading it for default execution rights.