The Day the Capability Race Built the Brakes: OpenAI’s 35-Second Emergency Stop

The story the AI industry tells itself is changing. “How strong can we build it” is no longer the only question; “how fast can we stop it” has joined it. The number that best illustrates this change is 35 seconds. OpenAI has made public the Control Room, a system that stops the output of its own frontier model within 35 seconds when a safety or security risk is detected. A company that sells models has produced, like a product, the very device that cuts off its models’ output. The party selling the most intelligent thing built the brake first. That is the core running through today’s news. While the capability race ran toward the limit, the control race quietly started on the other side of the same track. This post takes the number 35 seconds as its starting point and follows how “stopping” is becoming infrastructure for the entire industry. And at the end, we look together at what enterprises need on the ground.

An illustration of a brake clamping one point of a flowing data stream and stopping it there.

The 35-Second Button OpenAI Built

The Control Room is a centralized monitoring and shutdown device. It keeps watching research workloads and, when it sees a danger signal, stops that output within 35 seconds. What matters here is the word “centralized.” When there are hundreds of models and thousands of workloads, it is a structure that watches all of it from one place and turns it off in one action. That is a different scale from putting a button on each individual model. The speed of 35 seconds is equally meaningful. In the past, response came after an incident. The order was investigate, apologize, fix it in the next version. Now, the moment a signal appears, you must be able to turn output off and on again within seconds. The stop decision has moved from “an apology after the incident” to “a device that travels alongside.” Even for one company’s frontier model, running this structure means binding monitoring, judgment, stopping, resuming, and the record of that history together into a single system. In the end, safety goes beyond one line on a feature list, to the point where the architecture is designed with safety as a premise. The moment you leave a record of “who, when, and why stopped,” safety becomes the way of operating itself. A structure where the act of pressing the brake is itself an audit history. That is what the Control Room points to. The number 35 seconds therefore reads like a new unit for measuring safety level. Tomorrow’s models will be evaluated not only by how strong they are, but by how fast they can be stopped. How quickly and transparently you can stop. That is becoming the new competitive edge.

flowchart LR
    A[Research workloads<br>hundreds of models, thousands of tasks] --> B[Central monitoring]
    B --> C{Risk signal decision}
    C -- none --> A
    C -- stop within 35s --> D[Output halted]
    D --> E[History record<br>who, when, why it stopped]
    E --> F[Resume decision]
    F --> A

The structure where the Control Room binds monitoring, decision, stop, resume, and history into one system.

The Accelerator and the Brake in One Hand

This paradox repeats across one page of today’s news. Turn your gaze to the capability side and the frontier keeps climbing. Anthropic’s new model Claude Fable 5.2 designed three working electronic devices from a single-paragraph prompt. What stands out is that it finished three in one day. That included a home voice assistant hub, and on major benchmarks it outperformed GPT-6 Astra. The model has moved past text into physical design, things like circuit boards. Robotics company Skild AI open-sourced its physical solver Steel. It is said to be 13 times faster than PhysX. That means the bottleneck of virtual training environments has eased to that degree, and it is reasonable to expect simulation learning for agents and robots to speed up by the same margin. Capability is certainly growing, speeding up, and spreading into the physical world. But in news from the same time window, numbers in the opposite direction appear alongside. The cost of building model capability at Anthropic rose eight-fold over two years, and the capability improvement obtained per $1 billion actually shrank. The growing criticism around the 2x R&D baseline is in the same context. The price of gaining one inch of capability has grown that much. The harder you press the accelerator, the more expensive it gets, and the more expensive the capability, the more it is worth only if you can stop it. That is the moment the brake settles in as part of the product, rather than as an accessory. The more you sell expensive capability, the more you must sell alongside it the trust that you are someone who can stop it. The presence of a brake is becoming a standard specification beyond a differentiating feature.

35 Seconds and 95%: The Two Numbers of Stopping

Two numbers sit side by side, governing the stop. One is 35 seconds, the other is 95%. The 35 seconds is the time a company takes to cut off its own model. The 95% is the approval rate a city demands from its voters to lift a ban it itself put in place. The Bloomington, Indiana city council passed ordinance 2026-29, banning the use and deployment of superintelligent AI within the locality. Lifting this ban requires the approval of 95% of voters. The intent is to pin the unlock threshold as high as possible. If a tech company builds a “stop fast” device in 35 seconds, a local government builds a “hard to release” device in 95%. The directions differ. One is speed, the other is threshold. Yet both answer the same question. That stopping is a condition designed before execution, and it must be prepared in advance. The fact that 35 seconds and 95% run together in the news of the same morning is itself a signal that stopping has become the industry’s common language. Speed and threshold, two ways, have appeared together. The number 95% is especially noteworthy. Unlike ordinary decisions that only need to clear 50%, 95% is a threshold that effectively demands unanimity. Lifting a stop has been made far harder than imposing one. It reads as the most conservative safety device prepared for an incident. These two numbers, in the end, pose the same question at enterprise sites as well.

Where the Brakes Are Heading: Money and Hardware

The need for stopping is spreading to where money moves and to the hardware layer. Volaren, founded by a team of former Goldman Sachs people, has shipped a personal AI agent. This agent plugs directly into retail brokerage accounts, analyzes the portfolio, and automates trades. The moment an agent touches an account where real money moves, the question changes. Beyond “can it do the job,” “can it be stopped” and “can it prove what it did” become a single problem. For an agent that automates trading, the audit history is trust itself. Because a single malfunction can move the money in an account. That is why, for a securities agent, the history of actions becomes a more valuable asset than the speed of actions. In the hardware direction, movement in the same direction is visible as well. Google has expanded its 8th-generation AI accelerator lineup to four kinds. Two of the new variants target secure computing and large-scale data processing. It is a lineup expansion that squarely faces NVIDIA, and at the same time it is adding, of its own accord, the layer of “chips that compute safely.” If a chip presupposes secure computing, stopping comes to rest on the nature of hardware, not only on buttons in software. Money, simulation, chips. These are layers where stopping is needed. The places that need stopping are widening to every layer of the industry. The shape of the brake differs in each layer. In front of an account it is permissions and audit, and on top of a chip it is an isolated secure domain. But the common point is clear. Only if you can stop what is moving can you trust what is moving. It also means that without a brake, you are not even given the qualification to press the accelerator.

What Enterprises Need Is Not a Button but a Floor

Now we should bring the question to enterprises. OpenAI’s 35-second button stops “the output of one company’s frontier model.” The enterprise situation is larger than that. Agents are moving here and there, every day, touching each other’s data and systems. You cannot cut all of that off with a single central button. So the question an enterprise should ask is not “can we nail down the largest model” but “what did each agent do, under whose authority, and can it prove it.” The one that answers this question is Paxis. Paxis is ThakiCloud’s Agent-Native Cloud, a formal product already shipped (v1.1 GA). It treats Skills, Tools, Policies, and Audit Logs as first-class resources. It manages agent autonomy (L0~L3) through governance, puts a policy gate in front of each action, and leaves that history in audit logs. Execution happens in an isolated sandbox, and tools and functions are connected through MCP connectors and the skill marketplace. In a sovereign, on-premise K8s (ai-platform) environment, it can run inside the company’s internal network. Per-task model selection (CostRouter) also decides which model is used for which task. If the 35-second button is a device that “cuts off one output,” this floor is a structure that “wraps every action at every moment in policy and audit.” At the point when “how fast can we stop” becomes the industry’s hot topic, what a single button cannot fill is exactly this floor. The higher the agent’s autonomy, the higher the resolution of the policy and audit that wraps its actions must also be. Take Volaren’s securities agent as an example and it becomes clear. An action that moves money in an account sits near the top of the autonomy scale, but a policy gate must stand in front of it. Transfer limits, permitted accounts, time windows, approvers. Those are exactly those conditions. These conditions form the policy, and the agent must pass that gate every time it acts. If it cannot pass, execution stops inside the isolated sandbox, and even the attempt remains in the audit log. If the 35-second button is “turning off model output,” this structure is “conforming each agent action to conditions.” It is a much closer, and much more frequent, form of control. This floor is, at the same time, an answer to the industry pain today’s news revealed: safe execution of agents, sovereignty, cost.

In Closing

35 seconds is just a number. But the fact that this number made the headline and that even a city got locked with 95% means the industry has started selling “strength” and “stopping” together. As capability keeps growing, the structure that wraps that capability must grow at the same speed. What will decide the next stage of the capability race is not a bigger model, but what floor those models are allowed to move on. It is also a season in which the side that built the brake first ends up going farther in the end. What today’s morning news said together is that running speed and the confidence that you can stop have, for the first time, come to compete with the same weight. The next benchmark will include items like “time to stop.” And to fill that item, enterprises must translate today’s 35 seconds into tomorrow’s standard. The brake is no longer an option. It is an entry condition.

References

This post was written by synthesizing the news below.