Grok 4.7 lands: xAI pushes to 2.1 trillion parameters
Elon Musk's xAI shipped Grok 4.7 on September 12, 2026 — a 2.1-trillion-parameter model, up 40% from Grok 4.6's 1.5T. Per Musk's own framing, it trades a little serving speed for higher token efficiency and "outperforms all models" on engineering and scientific reasoning, with supplementary training on SpaceX data to sharpen real-world technical tasks. It's the third Grok iteration in under two months (4.5 in July, 4.6 in August), a cadence clearly aimed at forcing rivals onto xAI's timeline. Independent evals of the 4.6 baseline put it ahead of GPT-5.6 Sol but behind Claude Opus 5 and Fable 5 — so the 2.1T leap is the bet that raw scale closes that gap.
Kimi K2.8 Preview goes fully live with a 1M context window
Moonshot AI rolled Kimi K2.8 Preview out to all Kimi Code users on September 11. It keeps the same Model ID (kimi-for-coding), so clients and third-party tools pick it up with zero config changes — a quiet "default upgrade" rather than a headline launch. The pitch: performance close to flagship K3, with much better thinking efficiency than K2.7 Code, selectable low / high / max thinking effort, and a 1M-token context window open to every membership tier (K2.7 Code was capped at 256K). The signal is less the benchmark and more the strategy: Moonshot is quietly making long-context, thinking-capable coding the default floor, compressing the gap between its flagship and its workhorse model to under two months.
25 Fields Medallists warn AI is "strip-mining" mathematics
On September 11, twenty-five Fields Medal winners — including Terence Tao, Peter Scholze, Maryna Viazovska, James Maynard and Maxim Kontsevich — published "A Severe Misalignment of AI in Mathematics." Their argument: as AI labs race to solve famous problems as benchmarks, they're turning proof-generation into "mass production of true/false statements" that displaces the slower human work of writeups, attribution and absorption into the canon. The declaration follows OpenAI's September 8 claim that ~10,000 agents cracked the Navier–Stokes Millennium Problem, and OpenAI's September 10 withdrawal of its Caltech Mathathon sponsorship after mathematicians accused labs of research misconduct. No company is named; the complaint is about incentive structures, not capability.
The frontier is splitting in two directions at once: bigger models (Grok 4.7) and wider, cheaper deployment (Kimi's 1M default), while the research community pushes back on how the race is being scored. For daily AI roundups and deeper breakdowns, visit AI Nexus Daily.
Top comments (0)