METR: Anthropic redacted much of the "most informative" evidence from METR's 100-page report, so readers "must rely on our judgment more than we consider ideal" (@metr_evals)
METR released a 100-page evaluation report but noted that critical supporting material was heavily redacted by Anthropic prior to publication. The evaluation group shared this disclosure publicly through its official account to highlight the restrictions placed on the findings. Because key data was withheld, the public cannot independently inspect the evidence behind the assessment. Readers are ultimately forced to trust the evaluators' judgment far more than the organization prefers.
Key points
Readers must rely on the evaluation group's judgment rather than verifying the actual evi…: Readers must rely on the evaluation group's judgment rather than verifying the actual evidence themselves.
Anthropic chose to redact what the evaluation group described as the most informative evi…: Anthropic chose to redact what the evaluation group described as the most informative evidence.
The transparency and independent utility of METR's safety findings were significantly com…: The transparency and independent utility of METR's safety findings were significantly compromised by the omissions.
Sources checked by our analysis pipeline: 1
Dina Bass / Bloomberg: |
Wall Street Journal: |
Emma Roth / The Verge: |
Djohnson / CyberScoop: |
Andrew E. Freedman / Tom's Hardware: |
Jaron Schneider / PetaPixel: |
Davey Alba / Bloomberg: |
Tom Warren / The Verge: |
Tom Warren / The Verge: |
Radhika Rajkumar / ZDNET: |
Takashi Mochizuki / Bloomberg: |
Andrew Webster / The Verge: |
Brody Ford / Bloomberg: |
Meghan Bobrowsky / Wall Street Journal: |
New York Times: |
Tiffany Li / The Information: |
Jay Peters / The Verge: |
Jonathan Stempel / Reuters: |
Takashi Mochizuki / Bloomberg: |
Isabelle Bousquette / Wall Street Journal: |
Kelly Cloonan / Wall Street Journal: |
Carl Franzen / VentureBeat: |
OpenAI: |
Los Angeles Times: |
Ivan Mehta / TechCrunch: |
Ben Schoon / 9to5Google: |
Marcus Mendes / 9to5Mac: |
Reuters: |
The Information: |
Rya Jetha / Business Insider: |
This is a Techmeme archive page. It shows how the site appeared at 4:30 PM ET, November 5, 2026.
Transitioning software to power-efficient infrastructure often introduces complex compatibility hurdles, particularly when dealing with legacy architecture settings and unsupported native binaries. Developers can now utilize an automated assistant to evaluate their codebases, resolve blockers, and generate customized transition roadmaps. This support aims to accelerate transitions toward cheaper compute environments without requiring manual rewrites of core project configurations.
The newly launched weights equip autonomous vehicle engineers and roboticists with post-trained systems to generate synthetic multi-sensor environments and validate physical policies. Released under an open license permitting commercial deployment, these systems help researchers bypass traditional data collection bottlenecks by translating control inputs directly into synchronized, reactive video feeds.
The managed tool allows AI developers to instantly query external websites and receive fully synthesized, attributed answers without maintaining bespoke retrieval architectures. Operating via an AgentCore Gateway, the feature executes multi-step web searches, cross-references sources, and appends citations directly to natural-language outputs. This turnkey solution eliminates the infrastructure overhead of managing raw web scraping and synthesis pipelines.