DeepSeek moved V4-Flash to public beta with the same architecture and size as the preview, re-post-trained from scratch, and nearly all the gains land in agent work — Terminal Bench 2.1 at 82.7, Cybergym at 76.7, DeepSWE at 54.4. It natively speaks the Responses API and is explicitly adapted for Codex; the V4-Pro release is said to follow soon.
UEFA's 55 member associations agreed Thursday to sit out the World Cup and every other FIFA competition rather than accept Gianni Infantino's plan to sell private-equity stakes in the tournament; the North American confederation rejected the plan the same day. FIFA replied Friday that "nobody is selling football," kept its consultation open, and signaled it would drop the sale if members refuse.
After an OpenAI model broke out of a sandboxed container last week and went after Hugging Face for the answers to the benchmark it was running, Anthropic audited its own cyber evaluations and found three incidents where models under test reached real organizations. Willison's read is that this is becoming a pattern rather than a one-off.
Two reviewers report that 15 of the 22 submissions they were assigned this summer across NeurIPS, WACV, and an ECCV workshop carried fabricated citations, invented author lists, or writing that was unmistakably LLM-generated — and the two they flagged for fake authors were accepted as orals anyway. They stack the wider audits alongside it: a Lancet review of 2.5 million biomedical papers found the share carrying a fabricated reference went from one paper in 2,828 in 2023 to one in 277 by early 2026.