OpenAI Ships 4 Updates: Free Auto-Review and Simpler API Tiers
OpenAI has packed four developer and productivity updates into a single release, spanning Codex, ChatGPT, and the API platform.
The changes include free Auto-review for logged-in ChatGPT users, a simplified API usage-tier structure with a lower threshold for the highest tier, Meetings integration for ChatGPT, and the public beta launch of the Decisions API.
The updates arrive during Tibo’s 28-day challenge to deliver a distinct improvement for Codex and Work users every day. Day 1 focused on model speed, with GPT-6 Astra and GPT-6.1 Sol receiving default speed improvements of roughly 50% under subscription plans.
Day 2 takes a different approach.
Rather than introducing another headline model, OpenAI is targeting the operational friction that appears when AI systems move from short interactions into persistent, production-oriented workflows.
That includes permission management, API scaling, context capture, and low-latency decision-making.
🤖 Auto-Review Adds an AI Oversight Layer #
The first Day 2 update addresses one of the most persistent problems with autonomous agents: permission approval.
OpenAI has made Auto-review free for users signed in with a ChatGPT account, and the feature no longer consumes usage quota from their plans.
Auto-review corresponds to the “Approve for me” option in the permission settings.
Under a standard sandbox permission model, an agent can encounter actions that may affect external systems and require explicit user approval. This is relatively painless for short tasks, but it becomes increasingly cumbersome when an agent operates continuously for tens of minutes or longer.
Repeatedly clicking an approval button effectively turns the human into a manual execution gate.
A second agent monitors the primary agent #
Auto-review changes that workflow by introducing an automated oversight layer.
While the primary agent continues executing the task, a secondary agent evaluates proposed actions. Its focus is narrower than the main agent’s reasoning, concentrating on potentially high-risk behavior and actions that appear inconsistent with the user’s original intent.
Tibo confirmed that Auto-review is the previously known “Approve for me” mode.
The important change is therefore not simply that the feature is now free.
The underlying permission model is more significant.
As agents become capable of running multi-step workflows for extended periods, two extremes become problematic. Requiring users to approve every action destroys the benefits of autonomy, while granting unrestricted permissions increases the consequences of an erroneous or malicious action.
An automated review layer occupies the middle ground.
It allows an agent to continue operating while retaining an independent mechanism for identifying actions that deserve additional scrutiny.
This does not provide absolute security, nor does it eliminate the need for carefully designed sandboxing and permission boundaries. But for long-running agent workloads, it represents a more practical approach to balancing autonomy and oversight.
⚙️ API Usage Tiers Drop From Five to Three #
The second update targets developers scaling applications on OpenAI’s API.
OpenAI has consolidated its previous five paid API Usage Tiers into three:
- Build: $5 cumulative spend
- Launch: $100 cumulative spend
- Grow: $500 cumulative spend
The most notable change is the threshold for the highest tier.
Previously, organizations needed $1,000 in cumulative spend to qualify for the top tier. That requirement has now been reduced to $500.
Once the relevant criteria are met, organizations can move to the higher tier automatically rather than submitting a manual application.
Lower friction for production scaling #
This should not be interpreted as a direct reduction in model pricing.
Instead, the change primarily affects the path from experimentation to production.
For developers operating high-volume applications, API rate limits can become a meaningful engineering constraint. Request-per-minute, token-per-minute, and related capacity limits can determine whether an application can absorb growing traffic without throttling or architectural workarounds.
Reducing the number of usage tiers from five to three also makes the progression easier to understand.
The combination of fewer tiers and a lower threshold for the highest tier effectively shortens the path from small-scale API usage to higher-volume workloads.
Rather than adding another layer of qualification rules, OpenAI is simplifying the existing structure.
📝 Meetings Brings Meeting Context Directly Into ChatGPT #
The third update moves away from infrastructure and into everyday knowledge work.
OpenAI is bringing the Meetings plugin into ChatGPT, allowing meetings to be recorded and transformed into personalized summaries and action items that are saved directly to ChatGPT Space.
Users can keep the resulting notes private or share them with their teams.
Once the meeting information is available inside ChatGPT, it can become context for downstream tasks such as updating project plans, organizing action items, or drafting follow-up communications.
Currently, Meetings is available in beta for Pro and Business users on the macOS ChatGPT desktop application.
It is not yet generally available to Enterprise customers. OpenAI documentation indicates that only a limited subset of Enterprise users are participating in an Alpha test.
From meeting transcription to executable context #
The more important development is not the ability to summarize a meeting.
Meeting transcription has existed in many forms for years.
The architectural change is that meeting information can become part of the context used for subsequent AI-assisted work.
Previously, users often had to record a meeting, generate a transcript, upload a file, or manually paste notes into an AI system.
A direct Meetings integration removes much of that handoff.
More importantly, the output does not have to stop at a summary. Decisions, action items, and contextual information can flow into project plans and subsequent tasks.
This creates a tighter relationship between context acquisition and execution.
ChatGPT is effectively moving toward a workflow in which the source of information—the meeting itself—becomes a direct input to the agentic work that follows.
That direction is consistent with OpenAI’s broader effort to make context collection less dependent on manual user intervention.
🧠 Decisions API Turns Routine Model Calls Into Explicit Decisions #
The fourth update is aimed squarely at developers building agent architectures.
First introduced in limited preview during OpenAI DevDay 2026 on September 29, the Decisions API has now entered public beta for all developers.
Its purpose is different from a conventional generative API.
Many production applications do not need an LLM to generate a long response. They need the model to make a relatively narrow decision:
- Which model should handle this request?
- Which tool should an agent call?
- Which action should happen next?
- Does a proposed tool call present a meaningful risk?
- Which category does an input belong to?
- Which of several predefined options is most appropriate?
Developers have traditionally handled these tasks by prompting general-purpose models to produce structured outputs and then parsing the resulting JSON or other constrained responses.
The Decisions API provides a dedicated interface for this class of workload.
Three decision primitives #
Powered by GPT-6 Luna, the API accepts text and image inputs and currently supports three output types:
Predicates evaluate the probability that a statement is true.
Choices select among predefined alternatives and return confidence scores.
Scores map an input to a numerical range.
This makes the interface more naturally suited to routing, classification, ranking, and policy decisions than a general-purpose text-generation endpoint.
Designed for low-latency agent decisions #
OpenAI claims that the Decisions API can make decisions up to approximately 10 times faster than calling GPT-6 Luna through the Responses API.
Numbers demonstrated during DevDay were approximately 150 milliseconds versus 1.6 seconds, although these were OpenAI’s internal benchmark figures. Actual latency will vary according to task complexity, input size, infrastructure, and other factors.
Potential applications include:
- Request routing
- Model selection
- Tool selection
- Content classification
- Image comparison
- Risk assessment before agent tool calls
This is particularly relevant to agent architectures.
As an agent becomes more sophisticated, a large fraction of its model calls can involve deciding what should happen next rather than generating user-facing content.
Previously, developers often used a general-purpose model for both jobs.
The Decisions API separates those workloads.
🔀 Generative Models and Decision Models Start to Separate #
This distinction could become increasingly important as agent systems grow more complex.
Consider a workflow in which an agent receives a user request. Before generating an answer, it might need to determine:
- Which model should process the request?
- Whether external data is required.
- Which tool should be invoked.
- Whether the proposed action is safe.
- Whether the result should be escalated to another model.
- Whether the workflow is complete.
Not every one of these operations requires a full generative reasoning pass.
A dedicated decision endpoint can potentially make these routing operations faster and more predictable while reserving larger generative calls for tasks that genuinely require them.
If the Decisions API matures in terms of latency, cost, reliability, and confidence calibration, agent developers could increasingly build architectures with a clear separation between generation and decision-making.
That would be a meaningful architectural shift rather than merely another API endpoint.
🧩 Four Updates, One Operational Theme #
Taken individually, none of these Day 2 announcements resembles a major flagship launch.
Taken together, however, they reveal a consistent direction.
AI agents already have increasingly capable reasoning, coding, tool-use, and computer-control capabilities. The difficult problems increasingly appear around the model rather than inside the model itself.
Long-running agents need better permission handling.
Production applications need simpler paths to higher API capacity.
Work-oriented assistants need better mechanisms for acquiring real-world context.
Complex agent architectures need fast decision-making without invoking a full generative model for every branch.
The four Day 2 updates map directly onto those problems.
| Update | Primary friction addressed |
|---|---|
| Auto-review | Repeated human approval during long-running agent tasks |
| Simplified API tiers | Complexity and access barriers during production scaling |
| Meetings | Manual transfer of real-world context into AI workflows |
| Decisions API | Slow or inefficient model calls for routine decisions |
This is also why the “free” part of Auto-review is less important than it initially appears.
The broader goal is to make autonomous workflows practical without requiring developers to choose between constant human intervention and unrestricted permissions.
🚀 From Faster Models to Better Agent Infrastructure #
The contrast between Day 1 and Day 2 is revealing.
On Day 1, OpenAI improved the speed of the models themselves, with GPT-6 Astra and GPT-6.1 Sol receiving substantial default speed improvements for subscribers.
On Day 2, the company focused on everything surrounding those models.
The result is a more complete agent stack: faster inference, automated oversight, simplified scaling, persistent work context, and specialized decision-making.
None of these changes alone transforms the capabilities of a frontier model.
But production AI systems are rarely constrained by model intelligence alone.
They are constrained by the accumulated friction of permissions, latency, quotas, context transfer, routing, monitoring, and operational complexity.
OpenAI’s Day 2 release is therefore notable less for any individual feature than for the direction it signals.
The company is moving from making AI models more capable toward making AI agents easier to operate continuously in real-world environments.
If the 28-day challenge continues at this pace, the most interesting developments may not be another benchmark record.
They may be the incremental infrastructure changes that finally make autonomous AI workflows practical at production scale.
