METR publishes independent Anthropic AI R&D assessment with most key evidence redacted
lxrjl · x · 2026-11-10
- METR released an independent assessment (dated October 2026, published Nov 5) of AI R&D acceleration at Anthropic, authored by Tom Cunningham, Nate Rush, Parker Whitfill, and Hjalmar Wijk.
- Transparency caveat: Anthropic exercised redaction rights, and METR notes that a majority of the most informative evidence supporting its conclusions was redacted, described as removing confidential info that could accelerate other AI developers. High-level conclusions are shareable, but much underlying evidence is not.
- The reposter reads Anthropic's RSP as implying Model 4 will likely be its first "highly capable" model (crossing the AI R&D threshold, e.g. fully substituting its research staff at competitive cost within a 5x factor), substantially riskier than publicly analyzed models — so expect external review of the next Risk Report and a public Model 4 risk analysis by late November.
More from Models
- If a model scores below Kimi K3, blame the benchmark, not the model — nitarshan · 2026-11-10
- pyannote speaker-diarization-3.1 trends on Hugging Face — pyannote · 2026-11-10
- Fable 5.2 draws every pixel via code — no diffusion, no 3D, no render engine — DimitrisPapail · 2026-11-10
- Vibe-coded Google Earth in JavaFX shows Gemini 3.8 Flash is the best free workhorse — AnahataASI · 2026-11-10
- Gemini 4 Carbon goes live in Google AI Studio, requires Ultra plan or API key — cedric_chee · 2026-11-10
- Cutting GPU power 42% costs nothing: RTX 3080 Ti runs Qwen 3.5 9B at full speed on 250W — Asthenia5 · 2026-11-10