Today's (September 22) AI rundown: Xiaomi open-sourced MiMo-V2.6, with the Pro version scoring 46 on the Artificial Analysis Intelligence Index — the top open-weight model; xAI's Grok 4.7 benchmark landed the same day with the same 46, its coding-agent score jumping to 56; the Canadian province of British Columbia is suing OpenAI over a school shooting, alleging its safety team flagged the shooter's risky conversations but never alerted police; and Linear published a retrospective on how it reworked CI for the AI-coding era.
Xiaomi open-sources MiMo-V2.6: Pro scores 46 on Intelligence Index, topping the open leaderboard
On September 21, Xiaomi announced the MiMo-V2.6 series on its official blog: two natively multimodal models, Pro and Flash, plus Distill-Qwen-9B, more than 7,000 RL task environments and the training code — all open-sourced.
MiMo-V2.6-Pro scored 46 (46.32 precisely) on the Artificial Analysis Intelligence Index, beating Kimi K3 and Qwen3.8 Max to become the number-one open-weight model; it still trails closed frontiers like Claude Fable 5.1 and GPT-6 Astra. The standout detail is the "6-day livestreamed RL training": Pro and Flash each completed 30 steps and roughly 750,000 trajectories, costing about $2.62M and $0.85M, lifting out-of-sample DeepSWE v1.1 scores by about 17 and 14 points respectively. API pricing is unchanged from V2.5, and a new Pro-UltraSpeed hosting tier offers up to 20x output speed.
Further reading: 小米发布 MiMo-V2.6：全模态双版本，把 Agent 执行层做进模型里 (/article/2026-xiaomi-released-mimo-v26-dual-versions-of-full-modality-integrating-the-agent-ex)
Grok 4.7 benchmark: 46 overall puts xAI in the top four labs, coding agent hits 56
xAI released Grok 4.7 on September 21, and Artificial Analysis immediately scored it 46 on the Intelligence Index — 2 points above Grok 4.6, putting xAI among the index's top four labs, though still 7 points behind Claude Fable 5.1 and GPT-6 at 53.
The highlight is coding: paired with xAI's own coding agent Grok Build, it scored 56 on the Coding Agent Index, up 9 from 4.6, beating GPT-5.6 Sol and ranking fourth in native-harness setups. Elon Musk claimed on X that xAI ranks third in agentic coding, behind only Anthropic and OpenAI. On long-horizon real-world professional tasks (AA-Briefcase) it reached 1657 Elo, 111 above the previous generation and close to Claude Opus 5; GDPval-AA hit 1695 Elo (+90). Pricing is unchanged at $2 input / $6 output per million tokens.
British Columbia sues OpenAI: safety team flagged the risky chats, but never called police
On September 21, the government of British Columbia, Canada, sued OpenAI and CEO Sam Altman in the U.S. District Court for the Northern District of California (first reported by Reuters). The case stems from the February school shooting in Tumbler Ridge, which killed 9 people, most of them children.
The complaint alleges that the shooter, Jesse Van Rootselaar, was flagged as high-risk by OpenAI's safety team after discussing gun violence in ChatGPT conversations — but the company never forwarded the warning to law enforcement. The province argues the tragedy was preventable, seeking damages covering school rebuilding and mental-health costs, plus a court order forcing OpenAI to overhaul how it handles potentially violent conversations. More than 30 victims' families had already sued in California federal court; an OpenAI spokesperson called it an "unspeakable tragedy" and pledged cooperation with law enforcement.
Further reading: 不列颠哥伦比亚省起诉 OpenAI：内部标记风险账号却未报警 (/article/2031-british-columbia-sues-openai-risk-account-flagged-internally-police-never-alerte)
Linear shares its CI rework: the faster AI codes, the more expensive verification gets
On September 21, Linear published an engineering-blog retrospective on its CI overhaul, titled "CI costs are high" — quoted verbatim from the issue assigned by CTO Tuomas Artman.
The core argument cuts deep: AI coding has made writing code exponentially faster, but every PR still has to pass CI, turning verification into the new bottleneck — driving up infrastructure costs while developers and agents sit idle waiting for feedback. After the rework, the test suite grew nearly 4x since the start of the year, yet PR wait times fell from over 6 minutes to just above 5, and unit-test runner time was halved. The playbook: upgrade infrastructure and tooling, optimize blocking jobs, cut redundant setup, and speed up test execution. The post notes these optimizations apply beyond TypeScript to other languages and toolchains.
This briefing includes 4 items, each verified against an official source from the last 24 hours (official blogs, official announcements, Artificial Analysis benchmarks, Reuters' first report). Four other leads were excluded: the Kimi browser extension (the "WebBridge rename" actually happened on September 17), GGUF support in transformers (a long-standing feature, not a new release), NVIDIA Nemotron 3.5 Lightning (released August 11), and Step 5 Preview (released September 18, outside the 24-hour window).
