Thibault Sottiaux, a developer at OpenAI, says the company’s internal AI tool Astra was probably its “biggest competitive advantage” during the period it wasn’t publicly available. Writing on X, he claimed that once his team started using Astra internally, productivity improved enough that some product plans were pulled forward by six months. That’s not an incremental tweak. Six months is a strategic leap in AI development timelines, where the gap between leading labs and everyone else is often measured in weeks, not quarters.
The claim lands against a sobering backdrop. A recent study by IAPS fellow Severin Field surveyed 25 researchers across OpenAI, Anthropic, Google DeepMind, and Meta. Twenty of them ranked the automation of AI research as one of the biggest risks in the field. Several milestones those researchers had flagged as warning signs have already been hit. Perhaps most striking: half the respondents expect the most powerful models to stay locked inside labs, never sold to the public. That aligns with Anthropic’s own disclosure that Claude now writes more than 80 percent of its production code.
Not everyone is convinced the productivity gains are real, particularly when it comes to AI systems improving themselves. Self-improvement loops have long been a flashpoint in AI safety debates—the fear being that once an AI can meaningfully accelerate its own development, human oversight struggles to keep pace. Sottiaux’s post doesn’t resolve that tension. If anything, it feeds it. A tool that pulls timelines forward by six months is exactly the kind of capability that makes researchers nervous about recursive self-improvement, even if the tool itself isn’t literally rewriting its own architecture.
The bigger question is who benefits from these internal gains. If the most capable models stay private, as half the surveyed researchers predict, then productivity leaps like the one Sottiaux describes could widen the gap between frontier labs and everyone else—universities, startups, open-source communities, regulators. That’s not just a competitive issue. It’s a governance problem. When the tools that accelerate AI development are themselves concentrated inside a handful of companies, the rest of the world is left guessing what’s coming next.
