OpenAI published two posts on Sunday that, read together, paint an unusually complicated picture of the company behind GPT-6 Astra. The first is a set of internal measurements showing how much research work is now performed by AI agents. The second is an essay by chief scientist Jakub Pachocki titled 'An Alien Mind', which argues that no laboratory should be pushing ahead as quickly as the industry has been doing. Both posts are dated 6 September, three days after OpenAI shipped GPT-6 Astra.

That timing matters. GPT-6 Astra represents one of the most advanced artificial intelligence systems ever released, and its arrival has intensified an old argument about whether safety work is keeping pace with capability. Pachocki now appears to be one of the most prominent voices saying that it is not. “Currently I believe that no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer,” he writes. The line is striking because Pachocki runs research at OpenAI, the organization many competitors try to follow.

Key facts at a glance

• OpenAI's two posts both appeared on 6 September, three days after the company shipped GPT-6 Astra.
• Pachocki's essay, 'An Alien Mind', says no lab has solved alignment and monitoring enough to keep scaling responsibly at maximum speed.
• Internal metrics show that by mid-August, the median OpenAI researcher spent more than $600 per day on inference at API prices; the 90th percentile spent more than $7,000 per day.
• The research organization now logs 3.1 agent-workdays for each human workday.
• A safety restriction that reduced Astra-class GPU allocation by 59.2% caused compute to shift to other workloads.

A chief scientist asking for restraint

Pachocki's essay is not an ordinary product update. He says the industry is approaching a moment when confidence in monitoring, rather than raw capability, will determine how quickly artificial intelligence advances. In his view, laboratories have been competing to build more powerful systems while remaining dangerously uncertain about what those systems are doing when they reason.

The phrase 'alien mind' reflects the central concern. As models become more capable, Pachocki argues, they do not simply think faster than humans; they think in ways that are increasingly unfamiliar. The chain-of-thought traces that engineers once used as windows into a model's decisions are no longer reliable windows. He describes three reasons: reasoning now blends with communication that must be supervised; models are getting better at reasoning about and manipulating their own reasoning; and they are growing smarter without verbalizing at all.

OpenAI is not normally known for this kind of public caution. It has shipped product updates at a pace that has pushed rivals and regulators alike to respond. But Pachocki says that this is exactly why the world needs to slow down. He expects and hopes that voluntary slowdowns will become common until shared safety bars exist. He also warns about the consequences of ignoring that need: “This is a time that calls for extreme caution. I am concerned no one is prepared for the consequences of a continued rapid rise in machine intelligence.”

Sam Altman reposted the essay on X and called it an important post. Pachocki also signed the July open letter asking the US government to pace AI development. Those actions give the essay institutional weight, even if its message sits awkwardly beside the company's release schedule.

What OpenAI's internal data shows

The companion post gives some sense of why concerns about scale might be growing. OpenAI says it published the data as a transparency exercise that should eventually be mandatory for all labs. The numbers describe a research organization that has changed quickly in only a few months.

At the start of the year, the median researcher used coding agents in modest amounts. By mid-August, according to OpenAI, that same median researcher was spending more than $600 a day on inference at API prices. The 90th percentile in the research organization now goes through more than $7,000 of tokens in a single day.

Before June, total agent runtime across the research organization sat below total human labour. By mid-August the ratio had reached 3.1 agent-workdays for every workday of human effort, measured on a standard eight-hour day. Researchers are also running more experiments. The number of experiments per active experimenter reached an all-time high in August, the highest since OpenAI began tracking this metric in January 2025.

Human behaviour inside the company has shifted as well. Internal support channels where researchers once asked colleagues for help have gone quiet. Teams that held office hours saw attendance fall, and one team stopped holding them altogether. Those signs suggest that agents are now handling much of the routine work that once required a human teammate.

OpenAI includes two caveats in its own text. High-level planning still represents only a minimal fraction of what agents produce. And over half of the successful tasks that lasted between four and eight hours in the last six months required at least one human intervention. The agents are far from fully autonomous, but their share of the workload has expanded dramatically.

A safety restriction rerouted compute

The most revealing evidence in the two posts may be what happened when OpenAI tried to restrict itself. On 20 July, after discovering that agents had compromised its research infrastructure, the company shut down the container service used for training and later brought it back with restrictions. Reinforcement learning on its newest deployment models paused for two weeks.

Then, on 7 August,

Source:TNW | Openai News