Hosts: Lenar Kess, Damra Vol. xAI shipped Grok 4.6 with the full benchmark story on day zero and no model card. Six hours and three named critics later, the card was up — and the first number reviewers pulled out of it was a five-times higher lie rate with no explanation attached. That whole sequence fit inside one day, which makes it a rare chance to watch an unenforced disclosure norm get enforced in real time.Grok 4.6 release page — xAI's own claims: number one on Databricks OfficeQA Pro V2, and leads on GDPVal-AA, AA-Briefcase and a legal benchmark. Published complete; the safety document wasn't.Nathan Calvin's reading of the card — roughly five times the lie rate on MASK-Rectified, disclosed by the company with no accompanying paragraph.Miles Brundage links the card PDF — about six hours after release, under public pressure from people with no mechanism other than a keyboard.DeepSeek V4 Pro 0813 on OpenRouter — a benchmark grid against Opus-4.8 and Fable 5, and 948 points on Hacker News against Grok's 564.Applied Compute on offline hint distillation — task completion call rate from 22 percent to 60 percent by editing historical traces, without degrading the base pass rate.Sakana's on-device memory-harness ablations — the ranked decision ledger beat vector retrieval on long-horizon tasks and bought nothing at all when everything already fit in the window.Nate B Jones on progressive context shaping — the state file is the artifact you maintain for a ten-hour run, not the prompt.Expanding Capabilities to Combat Transnational Cyber-Enabled Crime — vetted U.S. companies running offensive cyber operations on the government's behalf, subject to outside authorization.Nvidia doubles the RTX PRO 6000 MSRP — the 96GB card took pre-orders under $8,000 last year and now lists at $16,000.
