Daily Brief · 22 September 2026
xAI ships Grok 4.7: 2.1T parameters, same price as 4.6
After at least five public delays since late July, xAI released Grok 4.7 on Monday—its largest model yet at 2.1 trillion parameters (up from 1.5T in 4.6) and live immediately in the Grok app, Cursor, Grok Build, and the xAI API. Pricing holds at $2 / $6 per million input/output tokens, with a 500K context and configurable reasoning. xAI pitches longer problem-solving, more self-checking, and supplemental SpaceX engineering data (Starlink telemetry, manufacturing, failure logs). Independent scores still place it second to Claude Fable 5.1 on GDPval and AA-Briefcase, and second to GPT-6 Astra on EEBench—so the launch is a capacity and distribution win more than a clear frontier claim.
What else matters
GitHub Copilot rolls out Grok 4.7 across IDEs
GitHub is adding Grok 4.7 to Copilot Pro through Enterprise SKUs—VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, and the cloud agent—at provider list pricing under usage-based billing. Business/Enterprise admins control access via model policy; defaults auto-enable new models unless explicitly disabled. Rollout is gradual.
Xiaomi open-sources MiMo-V2.6 Pro and Flash
Xiaomi published omnimodal MiMo-V2.6-Pro and Flash weights (text/image/video/audio, claimed 1M context) plus RL code and a technical report. Pro is a 1.02T sparse MoE with 42B active; API pricing stays at MiMo-V2.5 levels, with an UltraSpeed hosted tier claiming up to 20× output speed. Weights are on Hugging Face; OpenRouter also lists hosted access.
Light Origins launches Light-O1 for video-to-robot actions
Roger Jiang’s Light Origins released Light-O1, pretrained on structured human actions from internet video, then adapted to LightBot and Unitree G1 (kneeling, balancing, wiping, picking). An Apache-2.0 6B text-to-motion preview (Qwen3.5-4B-Base lineage) is on Hugging Face with a GitHub inference repo—testing whether web video can cut expensive robot-hour data collection.
White Circle opens Halo, its HF post-training stack
Paris safety startup White Circle released Halo, claiming 2.3–2.8× Hugging Face TRL throughput on gpt-oss-20b with lower peak GPU memory, keeping HF checkpoint formats. License is Apache 2.0 plus commercial terms: SaaS trainers over $20M related revenue need a deal; orgs over $20M ARR must credit “Trained with Halo.” Repo was early at launch (~30 stars).
