Anthropic AI Models Double Rate of Progress as Internal Automation Surges
A new investigation by METR reveals Anthropic is using internal AI models to accelerate its own development, yielding two years of progress in twelve months.
Frontier AI lab Anthropic is leveraging its own artificial intelligence models to speed up research and development, cutting normal progress timelines roughly in half. According to an evaluation released by AI safety research organization METR following an 11-week probe inside the company, Anthropic’s models are advancing at approximately twice their speed from a year ago. Standardized benchmarks suggest the company achieved the equivalent of two years of historical technical progress over the past twelve months alone.
Crucially, METR reported high confidence that this rapid improvement cannot simply be credited to scaling computational resources. Instead, researchers concluded that the majority of the acceleration observed as of October stems from AI tools directly assisting Anthropic's staff, with that contribution steadily rising. While human researchers still guide the process, METR noted early indications that fully autonomous contributions may expand soon, finding no visible bottlenecks or diminishing returns that would halt further speedups.
Internal models and massive productivity leaps
The speedup has been propelled by internal models that have not been deployed publicly. According to METR, Anthropic finished training three unreleased models over the past six months: designated under the temporary codenames Model 2 (completed in July), Model 3 (completed August–September), and Model 4 (completed September–October). The evaluation also noted an earlier system, dubbed Unreleased 2022.
Anthropic employees working with the lab's most advanced internal systems self-reported being roughly 12 times more productive than they would be without AI assistance, based on internal Slack polls and interviews. That figure represents approximately double the productivity multiplier employees reported just months prior while utilizing Mythos 5.
Backlash over heavy redactions and safety
While Anthropic granted METR unusual internal access—sharing non-public data, model metrics, and staff access—the release of the findings has triggered significant industry debate. Anthropic exercised its contractual rights to heavily redact the public version of METR’s roughly 100-page report, removing what METR described as a majority of the most informative evidence.
Anthropic co-founder Sam McCandlish defended the redactions on X, arguing that detailing the precise drivers of the acceleration would be commercially sensitive and could provide competitors with an undesirable roadmap to speed up. Even so, the secrecy has drawn scrutiny. Daniel Kokotajlo, a former OpenAI researcher who directs the AI Futures Project, sharply criticized the lab for shielding details while continuing to accelerate capability work. Brianna Rosen of the Institute for AI Policy and Strategy similarly stressed that democratic oversight requires independent evaluators and governments to verify evidence directly.
External safety observers, including Apollo Research CEO Marius Hobbhahn, indicated that the reported acceleration matches or exceeds timelines outlined in scenarios like AI 2027. Amid concerns regarding recursive self-improvement (RSI), New York state Assemblyman Alex Bores reacted to the report by urging Congress to pursue an immediate legislative ban on RSI.
What it means for developers
For software engineers and technical teams, the METR report signals that the capabilities of frontier foundation models are shifting much faster than historical training runs predicted. Alan Chan, a senior research fellow at GovAI, noted that the findings could compress industry timelines toward systems capable of fully automated software engineering tasks.
As AI labs employ internal models to automate their own coding, tooling, and infrastructure optimization, developers outside frontier labs will need to adapt to shorter release cadences and rapid capability shifts. Development workflows centered around manual prompt crafting and static scripting will likely need to make room for semi-autonomous agentic pipelines that emulate how labs are operating internally.
For engineering teams seeking to evaluate these rapid advances across providers without lock-in, developers can try top AI models cheaply through one API at https://apixoai.online. Tracking model evolutions side by side will become increasingly vital as frontier labs continue to deploy successor architectures trained with the help of automated machine intelligence.
Source: Why a new report on Anthropic automating AI development is alarming researchers — The Verge AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
