A developer at OpenAI says Astra has raised the company's internal productivity enough that part of its plans were brought forward by six months. That is a claim about a calendar rather than a benchmark, and the calendar is the harder of the two to audit: there is no score to reproduce and no eval to rerun, only a company reporting that its own software made it faster at building its own software. It is also, in outline, the threshold the field has spent years describing as the one that changes the shape of everything after it.
Recent work by Severin Field, a fellow at IAPS, suggests the view inside OpenAI is not an outlier. Field surveyed 25 researchers at OpenAI, Anthropic, Google DeepMind and Meta. Twenty of them named the automation of AI research itself as one of the largest risks the technology carries. Several of the thresholds those researchers pointed to have already been passed.
Anthropic, separately, says Claude now writes more than 80% of its own code for production systems.
Set those two statements side by side and they describe the same loop from opposite ends. One lab reports that the loop has started tightening enough to move dates on a roadmap. Another reports the input side of it: the code that builds the product is now mostly written by the product. Neither company is describing an experiment. Both are describing how the work already gets done.
Not everyone accepts the productivity estimates, and the disagreement is sharpest exactly where the stakes are highest — on whether models are meaningfully improving themselves. That skepticism seems to me well placed, for a structural reason rather than a suspicious one: every number in this story is produced, measured and published by the party it flatters. A share of code is the most generous available unit. It counts lines, not decisions, and nothing in the figure says which code, or who chose what to build. "Part of the plans moved by six months" is looser still. Moved against which baseline, set when, and judged by whom? No one outside the company can check, and nothing in the claim invites them to.
What the claim is quiet about is more interesting than what it asserts. A six-month schedule change is the kind of internal fact that a lab would normally keep to itself; stating it publicly turns an unverifiable operational detail into a capability claim, and the capability claim is the part that reaches investors and regulators. That is not a reason to assume it is false. It is a reason to notice that the only evidence offered for the most consequential kind of AI progress is a company's own account of its own schedule.
There is a second finding in Field's survey that fits uncomfortably against the first. Half of the participants expect the most capable models to stay inside the companies that build them and never be sold to the public at all. If both hold — acceleration driven by internal models, and internal models that never ship — then the public record of AI progress stops being the models and becomes the release dates. The frontier does its work behind the door, and the only trace it leaves outside is things arriving earlier than anyone expected, with no way to establish why.