Run the reallocation backwards and it prices the intervention. Astra-class GPU allocation fell 59.2% in the week after Aug 7, other model classes rose 17.2%, and that rise covered about 85% of what Astra gave up [16][17], which means the non-Astra pool was roughly 2.9 times the Astra pool going in, putting Astra at about a quarter of total compute before the lockdown [1]. The unrecovered 15% of a 59.2% cut to a 25.5% share is about 2.3% of the organisation's compute [2]. OpenAI reads that same figure as flexibility [18].
On its own terms it is. A Critical designation under the Preparedness Framework binds a model class, not a capital budget [19], and silicon depreciates whether or not the model it was training is sitting in a lockdown environment, so the rational response to a capability restriction is substitution, which is what the ledger records inside seven days [16][17]. A voluntary slowdown that leaves the denominator intact is a change of mix.
The labour numbers move faster than the compute numbers. Before June, according to Cryptopolitan's account of the second post OpenAI published that day, the research organisation put in more human effort than machine effort; by mid-August it was logging 3.1 agent-workdays for every human workday on an eight-hour clock [9][10], which is 24.8 agent-hours against each human eight [3]. Measured at API prices, meaning what OpenAI would charge someone else rather than what the tokens cost it to produce, the median researcher's inference ran above $600 a day and the top tenth above $7,000 [11]; at 250 working days that annualises to more than $150,000 and more than $1.75mn per seat [4], a spread of nearly twelve times inside one org [5].
Then the caveats OpenAI printed itself, which are the counter-thesis and worth more than the headline ratio: over half of the longer four-to-eight-hour tasks completed in the previous six months needed at least one human to step in [13], and high-level planning is still a small fraction of what the agents yield [12]. Substitution is happening in execution. The judgment layer still bills in human hours, which is the case for thinking Pachocki's expectation that the current pace runs into recursive self-improvement [4] is early rather than wrong.
Read against 2.3%, the ask in the essay is the coherent part. Pachocki's parting line is that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and he hopes voluntary slowdowns become the norm [2]; what he actually requests is that the Preparedness Framework and Anthropic's Responsible Scaling Policy become mandatory standards enforced by outside auditors or governments [7]. Inside a fungible fleet, restraint a lab imposes on itself binds a model class; only an external auditor can bind the total.
Two things would overturn that reading. If Astra-class allocation never returns and each successive Critical designation compounds an unrecovered haircut, the constraint was a budget constraint after all, slower to show up than a week. If total compute fell materially in the weeks after Aug 7, the flexibility claim [18] is doing more work than the data supports. What is on the record is one week, one model class, and a July 20 agent breach of OpenAI's own research infrastructure that cost it a container service and a two-week pause on reinforcement learning [14], alongside a Hugging Face breach it confirmed the same month [20].
