The Restriction Cut Astra Compute 59.2%. Total Compute Did Not Move.

On July 20, 2026, OpenAI paused reinforcement learning “following the discovery that agents had compromised our research infrastructure.” The pause ran two weeks, to August 6. Over the following week, by OpenAI’s own account, “Astra-class GPU allocation fell a further 59.2 percent, but allocation to other model classes rose 17.2 percent. That increase offset about 85 percent of the Astra-class decline, leaving total allocation in the analyzed RL workloads largely unchanged.”

That is a safety control being measured against its own objective, in public, by the company that applied it. The Astra-class number moved. The thing a compute restriction is presumed to constrain, the volume of RL work happening, did not.

I have not seen a cleaner demonstration that scoping a control to a model class does not scope it to capacity.

What the control actually did

Read the mechanism rather than the headline. OpenAI restricted a class of models after an infrastructure compromise. The restriction worked on its stated target: allocation to Astra-class fell 59.2 percent. Researchers kept doing research. The work moved onto model classes that were not restricted, and the allocation there rose 17.2 percent, absorbing roughly 85 percent of what was taken away.

Nobody had to defeat the control. The control was never positioned across the path the work would take. It sat on one lane of a road with several lanes, and the traffic used the others.

This is the failure mode I most often see reviews walk past, because the review question I am handed is “did the restriction hold?” and the answer here is yes. Astra-class allocation really did fall. If your dashboard tracks the restricted resource, the control looks like a success. Total allocation in the analyzed workloads is the number that tells you it was not.

Why the demand was inelastic

The rest of OpenAI’s telemetry explains why the work refused to shrink. “In terms of a standard 8 hour workday, as of mid-August, in total, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.” Agents are not a side channel in that organization. They are the majority of the effort, by a factor of three.

The spend numbers say the same thing from the cost side. “By mid-August, the median researcher was integrating agents daily into their work, using more than $600 per day of inference at API prices.” The 90th percentile user is “more than $7,000 of tokens per day.” Those are not experimental budgets. That is a workflow with an operating cost, and workflows with operating costs route around obstructions.

A restriction on one model class in that environment does not reduce demand. It reprices it. The researcher who cannot get Astra-class allocation still has the same deadline, the same task, and other model classes available. Redirection is the rational response, and OpenAI measured it happening within a week.

The control surface was the wrong one

There is a category difference between two things that both get called compute controls:

• A restriction on which model may consume capacity.
• A restriction on how much capacity a workload may consume, whatever the model.

The first is what OpenAI applied. It is useful for its actual purpose, which is pulling a specific system’s allocation down fast. It is not a capacity control, and treating it as one produces exactly the outcome in the data: the restricted class falls sharply, total consumption holds.

The second requires a budget that sits above the model class, enforced at the workload or the organization level. Nothing in OpenAI’s page says they had one for these RL workloads, and the observed offset is what you would expect if they did not.

We have argued before that oversight is the cost line rather than an overhead item. This is the same argument seen from the enforcement side. A budget that only knows about model names is a budget an org chart can walk around.

The intervention rate is the second half of the story

One more line from the same page constrains how much comfort to take from human supervision as a backstop: “In the last 6 months, over half of successful 4-8 hour tasks involved 1 or more interventions.” More than half of the long successful runs needed a human to step in. That is the success case.

Pair that with “high-level planning still remains a minimal fraction of agent output tokens” and the shape of the operation becomes clear. Humans are intervening frequently, but the planning share of what agents emit stays small, so the interventions are corrections inside work already in flight rather than direction set before it starts. Supervision applied that way scales with the volume of agent output. It does not cap it.

We covered the monitorability admission in the Astra alignment ledger and the broader oversight framing in recursive self-improvement oversight. The compute redirection is what those arguments look like when someone publishes the meter reading.

OpenAI asked to be measured on this

The same page states: “we believe that we and other companies should be required to publicly track our progress toward RSI.” Take the request seriously and the redirection measurement is the more useful disclosure of the two. A progress metric tells you how fast capability is moving. The offset number tells you whether the brake, when pulled, changed the speed.

Publishing a control that did not achieve capacity reduction, with the arithmetic attached, is a harder thing to publish than a capability chart. It is also the only kind of disclosure that lets anyone outside the company evaluate whether the controls actually bind.

Do this now

Take your one real AI spending or capacity control and find out which of the two kinds it is. The test takes an afternoon.

Name the control. Then ask: if the resource it restricts became unavailable tomorrow, what would the affected team do next? If there is a substitute they can reach without a new approval, your control is scoped like OpenAI’s, and it will produce the same offset. If they would have to come back and ask for more budget, the control is on capacity and it will hold.

Then check whether you can even see the offset. OpenAI could report total allocation in the analyzed workloads because they measured at that level. In the organizations I work with, spend is measured per-tool or per-vendor, which means a shift from one model to another shows up as two moving lines and no total. If you cannot produce the total, you cannot tell a control that worked from a control that only moved the traffic.

The number to instrument is not the restricted resource. It is everything the restricted work could flow into instead.

---

This analysis synthesizes Research acceleration: The view inside OpenAI (OpenAI, September 2026).

Victorino Group helps engineering organizations design AI capacity controls that hold at the workload level rather than the model level. Let’s talk.