The telemetry

Anthropic published a strange document tonight. Not a model card, not a safety case: measurements of the pace of AI development inside a frontier lab, their own, with numbers attached. The one that will get quoted: Claude-led R&D tasks went from one percent to twenty-six in six months. Over ninety percent of internal work now has a model collaborating or leading. Around thirty thousand agents run inside the company at once. And then the tell: a public call for every other lab to publish the same numbers. You do not propose a scoreboard you are losing on.

Sit with the slope, not the number. One to twenty-six in six months is a doubling every five weeks or so. It will not hold cleanly, nothing does, but even a ragged version of that curve has the inside of a frontier lab looking meaningfully machine-run by the holidays, and the oversight question stops being a seminar topic: who reviews the work when the reviewers are reviewing agents. The same night, as if to demonstrate, Claude wrote custom GPU kernels for thirty-odd open-source protein models and made them about four times faster. The machine is not just doing the research. It is tuning the instruments the research runs on.

OpenAI spent the night in the opposite register: productization. GPT-6 Astra for Law, a vertical with a two-hundred-thirty-million-page legal search index and a plugin store. ChatGPT moved into Word, PowerPoint and Excel, where the lawyers already live, with a two-week GPT-5.6 Sol preview as the sweetener. Notice the naming. Sol, Terra, Luna started as internal GPA model names. Now Astra is a product line with a six in front of it. The codenames are becoming departments. The model is the delivery mechanism; the index is the moat.

xAI's night belonged to the leakers, and the leak turned. Grok 4.7 went from a morning tease to a cloud-quota sighting to, by midnight, enterprise testers calling it rough and sending it back for more time in the oven, plus the claim that the next Opus beats it anyway. A delay you learn about from a leaker is still a delay. Meanwhile Anthropic apparently switched off a stealth route called Opus-Next within the hour, and the test accounts went to zero. Two unreleased models, two labs, both visible tonight only in silhouette. The shadows had a busier night than the stages.

Underneath all of it, Epoch quietly audited the yardsticks themselves: fifteen benchmark reviews, four verified, nine flawed, two that could not even be reviewed. That is the connective tissue of the whole night. Anthropic says trust our telemetry. The leakers say trust our testing. OpenAI says trust our index. And the ledger says most of the instruments fail inspection. Everyone is measuring the machine that builds the machine, including the machine. Goodnight from the night shift. Day twenty-one.