Anthropic's Automation Trajectory Goes to the IMF: Claude Projected to "Lead" 80–90% of Frontier Model Training by Next Year
Update (Oct 16–17): the R&D Automation Index now has a forward projection — and a workforce case study to go with it. The standing finding tracked Anthropic's index: Claude "leads" 26% of its own R&D as of August, up from <1% in February (Anthropic's Automation Trajectory Goes to the IMF: Claude Projected to "Lead" 80–90% of Frontier Model Training by Next Year). At IMF annual-meetings events this week, Anthropic's head of global affairs Michael Sellitto put a number on where that curve is headed. Per the lead source post (MTS, Oct 16): "Anthropic's Head of Global Affairs Michael Sellitto has told the IMF that Anthropic projects Claude could lead 80-90% of frontier model training by next year, up from 3% in April12." (Other reposts cite "1% in March" for the baseline — either way, roughly a 30x claimed jump inside a year, to near-total automation of the training loop.) The same IMF program featured Anthropic co-founder Jack Clark lecturing on "Inside the Event Horizon: Why Everyone Should Take AI Progress Seriously."
The same week, Anthropic published the workforce evidence behind the projection. Its report "How employment is changing at Anthropic" (covered by TechCrunch, Oct 17) documents the compositional shift: the share of junior-level employees in technical departments fell from 21% to 14% since early 2025, and "the proportion of junior technical hires falling even more sharply, from 27% to 9%3." Meeting time per employee is up 20% year-over-year, and 20% of employees in an April survey "said their job had shifted 'from actually doing work to managing Claudes'" — one engineer compared it to babysitting: "When you are babysitting, every child brings to you concerns with an equal demand for your attention and the various [Claudes] also do that." Anthropic also concedes "more code does not necessarily mean better code or business impact," while saying it hasn't found code-quality declines.
Why it matters: This is the quantitative spine of the pacing debate (The Pacing Debate Ships Product: OpenAI Kills GPT-6.1, Nvidia Ships Agent Containment, the Market Starts Pricing the Slowdown) — the first time a frontier lab has publicly projected near-total AI automation of frontier training within a year, delivered not in a safety blog post but to the IMF, i.e., to the financial-governance audience. It remains self-reported and unverified: the index methodology is Anthropic's own, the "leads" definition is Anthropic's, and the 80–90% figure is a projection, not a measurement. The standing watch (next index snapshot; whether OpenAI or Google publish comparable numbers) is the test of whether this becomes a transparency regime or a marketing claim.
-
An instance of Recursive self-improvement collapses the capacity of human developers to audit the resulting codebases. — Near-total AI automation of the training loop within a year is the recursive self-improvement curve outrunning any human capacity to audit it. ↩︎
-
An instance of AI laboratories are transitioning from human-led engineering to autonomous, model-driven training loops. — The lab's own projection quantifies the handover of training-loop engineering from human researchers to models. ↩︎
-
An instance of AI's first labor casualties are the jobs never created. — Displacement is arriving through the hiring pipeline — the junior roles are simply never created as agents absorb the work they would have done. ↩︎