Anthropic’s AI‑Powered Research Surge Raises Safety Concerns as Company Calls for Slower Development
A METR investigation finds Anthropic’s AI models are improving twice as fast as a year ago, with internal productivity gains of up to 12‑fold, while the firm heavily redacts the report and continues to push an IPO despite its CEO’s public plea to slow AI progress.
Anthropic’s AI research is accelerating dramatically
An 11‑week probe by the AI safety group METR, conducted inside Anthropic, shows that the company’s AI models are advancing roughly twice as quickly as they were a year ago. Benchmark data indicate “about two years of historical progress over the last year.” METR researchers say the speed‑up is not simply the result of more compute; instead, the majority of the acceleration in October was attributed to AI‑assisted work itself.
Productivity gains for Anthropic staff
Informal Slack polls and employee interviews reveal that researchers using Anthropic’s most capable internal model report being about 12 times more productive than without AI assistance. This figure is roughly double the productivity self‑reports from a few months earlier, when staff used the earlier internal system, Mythos 5.
New internal models driving the surge
METR identified three unreleased models that Anthropic completed training in rapid succession:
| Codename | Training period |
|---|---|
| Model 2 | July 2026 |
| Model 3 | August‑September 2026 |
| Model 4 | September‑October 2026 |
An additional “Unreleased 2022” model from 2022 was also mentioned. These models remain under wraps, and their capabilities are believed to underpin much of the reported productivity boost.
Heavy redactions limit public insight
Anthropic granted METR access to confidential research, performance data, and employee interviews—more than METR initially expected. However, the company exercised its right to redact a majority of the evidence in the public version of the report. METR notes that many of the omitted sections were “high‑level” but still essential for independent verification. Anthropic co‑founder Sam McCandlish defended the redactions on commercial grounds, arguing that disclosing the exact drivers of the speed‑up could give rivals a “roadmap” to accelerate their own work.
A paradox between rhetoric and practice
The findings arrive as Anthropic prepares for an IPO, while CEO Dario Amodei has publicly urged AI labs to “pump the brakes” on development. Critics see a contradiction: the company champions safety and calls for slower progress, yet its internal processes appear to be fast‑tracking AI capability growth.
Industry reaction and policy pressure
- Marius Hobbhahn, CEO of Apollo Research, says the report suggests AI‑enabled R&D is moving faster than many forecasts, potentially outpacing safety measures.
- Alan Chan of GovAI argues the results should make skeptics reassess timelines for fully autonomous AI contributions.
- New York Assemblyman Alex Bores called on Congress to ban recursive self‑improvement (RSI) “this year if possible,” reflecting growing legislative interest.
Outlook: Transparency and oversight in focus
The METR report has intensified calls for greater transparency and independent oversight of frontier AI labs. Researchers such as Brianna Rosen (Institute for AI Policy and Strategy) warn that without unrestricted access, public trust and democratic oversight remain fragile. Anthropic has announced plans to publish its own report on AI progress “in the coming weeks” and to outline new safety steps, but the extent of forthcoming detail remains uncertain.
Bottom line: Anthropic’s internal use of AI is dramatically accelerating its research pace, even as the company publicly urges caution. The heavy redaction of METR’s findings fuels debate over corporate responsibility, transparency, and the adequacy of existing safety frameworks in a rapidly advancing AI landscape.
