Report finds Anthropic’s AI models improving twice as fast with AI assistance
METR’s 11-week investigation inside Anthropic documents rapid internal progress and productivity gains, but warns of opacity in unreleased model details.
1 source · cross-referenced
- Anthropic’s internal AI models are improving roughly twice as fast as a year ago, with most acceleration attributed to AI assistance rather than compute scaling, according to a METR report after an 11-week investigation.
Anthropic’s internal AI models are improving roughly twice as fast as a year ago, with benchmark scores indicating ‘roughly two years of historical progress over the last year,’ according to a report by the AI safety research organization METR. The 11-week investigation, published Thursday, attributes most of this acceleration to AI assistance rather than increased compute power, a historically dominant driver of progress.
METR researchers said they are confident the speedup is not solely due to scaling up training compute, and that the contribution from AI has been growing over time. They also noted that Anthropic employees using the company’s most capable internal model reported being around 12 times as productive on average compared to working without AI assistance. This is roughly double the productivity gain reported just months earlier while using an earlier internal model.
The report highlights that Anthropic has completed training on three unreleased internal models—codenamed Model 2, Model 3, and Model 4—between July and October, reflecting a swift development pace. METR also identified an unreleased model from 2022, referred to as Unreleased 2022. Anthropic provided METR with access to non-public information, including confidential research and model performance data, though the company retained the ability to redact portions of the final report.
METR acknowledged Anthropic’s cooperation exceeded initial expectations but emphasized the limitations of its access. Anthropic could redact significant portions of the report, and the investigation’s pipeline of information was narrow. METR warned that the public version of the report lacks much of the evidence it found most informative due to redactions.
- Nov 10, 2026 · Wired
Kora, a Linux-based desktop OS from Cyanogen cofounder, targets AI power users with local model integration
Trust76 - Nov 10, 2026 · One Useful Thing — Ethan Mollick
Mollick argues new AI models approach "jagged AGI" across coding, business, and humanities
Trust78 - Nov 9, 2026 · The Verge — AI
Former DOJ antitrust chief launches AI legal-research startup to challenge monopolies
Trust72