The AI world moves fast and the hype moves faster. Every story here is checked against its sources before it goes up — exciting framing, verified facts, never fabricated. New stories land every weekday.
Archive: these are the stories we researched and wrote ourselves (September–November 2026). For today's headlines, see the AI news wire.
— Google's fast new Gemini 4 model is now the default brain in AI Mode and the hardest AI Overview questions — and developers can try it free in AI Studio.
Read the story →— Anthropic's Claude now joins Google Meet calls as a visible participant, taking notes and building docs and whiteboards live — but only for Team and Enterprise plans.
Read the story →— OpenAI's new Daybreak products send agent teams to attack your web app, then write the fix — with a human still approving every merge.
Read the story →— Thinking Machines Lab put its 594-billion-parameter Inkling 2 Small on Hugging Face under Apache 2.0 — free to download, run, and build on commercially.
Read the story →— Google's new decision model scores yes/no, multiple-choice, and rubric questions in a single pass — no text generation, no output-token bill.
Read the story →— DeepMind's newest robot brain ran Apptronik humanoids on active production lines in Hungary and Germany — but the public can't use it yet.
Read the story →— The 280-billion-parameter model runs just 14 billion parameters per token, with a million-token window — and Baidu says the weights go public under Apache 2.0 in two weeks.
Read the story →— The video-trained world-action model ships with full weights and fine-tuning code under Apache 2.0 — the most permissive robot model release of the week.
Read the story →— Google's strongest model is now available to paying developers and AI Ultra subscribers at $2/$10 per million tokens — but the introductory price doubles on December 1.
Read the story →— Meta's smallest Muse model matches Claude Haiku 5.5's intelligence score at about a third of the cost per task — built for the repetitive work AI agents do all day.
Read the story →— Cursor's multiplayer Projects beta lets teammates — and their coding agents — share context, a task board, and handoffs in one workspace.
Read the story →— Anthropic's new flagship models cost $8/$40 per million tokens — 20% less than Fable 5.1 — run up to 30% faster, and Mythos 5.2 stays limited to vetted research organizations.
Read the story →— NVIDIA unveiled GR00T 2 at the Conference on Robot Learning: a robot foundation model trained on 402,730 hours of human video that completed 40 of 50 new tasks with no task-specific demonstrations.
Read the story →— ByteDance's new image model claims up to 100 consecutive edits with pixel-stable untouched regions, plus native transparent output — but the benchmarks are its own and no API price or model ID is published yet.
Read the story →— Anthropic's new flagship performs at the level of the top-tier Fable 5.1 on most work at $4/$20 per million tokens — 40% less to run than Opus 5, with cache reads cut 60%.
Read the story →— OpenAI's new GPT-6 Sol and Luna cost half what their predecessors did — Luna starts at $0.10 per million input tokens, the cheapest model OpenAI has ever shipped.
Read the story →— Anthropic is partnering with OpenEvidence to give doctors in roughly 100 low- and middle-income countries free AI-powered clinical decision support, drawn from peer-reviewed research.
Read the story →— Alibaba CEO Eddie Wu announced on September 22 plans to train a 5–10 trillion-parameter model, unveiled the Zhenwu V900 AI chip — claimed to be China's most powerful — and set a 20-gigawatt data-center target for 2032.
Read the story →— Belgian cybersecurity firm Aikido released Altar on September 21 — an open-weight AI model for defensive security based on Z.AI's GLM-5.3, compressed to run fully on-prem (including air-gapped networks) so sensitive code never leaves the customer.
Read the story →— Xiaomi published ungated, MIT-licensed weights for MiMo-V2.6-Pro-RL and Flash-RL — a 1.02T-parameter sparse MoE with 1M-token context and omnimodal input, ranked the world's top open-weights model by Artificial Analysis.
Read the story →— Grok 4.7 arrived September 21 with a 500k context window, adjustable reasoning effort, and pricing held at Grok 4.6 levels — $2 per million input tokens and $6 per million output tokens. It is live in Cursor, Grok Build, and the Grok API.
Read the story →— StepFun’s Step 5 Preview is a 600B-parameter agent model with a 1M-token context at $2.70 per million output tokens — about one-seventh the price of comparable models — with open weights promised October 15.
Read the story →— xAI’s Grok Voice Transcribe 2.0 is live as the default speech-to-text API — claimed 2× the accuracy of 1.0, word-level timestamps, free diarization, and the same $0.10 per audio hour pricing.
Read the story →— Mark Zuckerberg opened Meta Muse connector development to third-party developers on September 18 — ten days after launch — kicking off the agent app-store land grab.
Read the story →— Meta's Muse, Instinct, xAI's Grok Bot, and OpenAI's GPT-6 Astra are the four personal AI agents that matter right now. Here's what each one does and which you can actually use today.
Read the story →— TypeSafe AI emerged from stealth on September 15, 2026, with a $40M seed led by DCVC and Jev — the first 'System One Model.' It's not a chatbot: it returns typed, probabilistic decisions for software in 70–500ms at $0.042 per million tokens, with output free.
Read the story →— Open models are going agentic: iFLYTEK's Spark-X2.5 (1.7B/4B, launched September 1, 2026) packs 1M-token context into on-device models, and NVIDIA's Nemotron releases — Content Safety 4B, streaming ASR, Lightning 30B-A3B, and the RTEB-topping Nemotron-3 embeddings — give builders self-hostable agent infrastructure.
Read the story →— i-have-adhd by ayghri is an installable skill that forces AI coding assistants to lead with the answer: action first, numbered steps, no preamble. It topped GitHub's weekly trending chart with ~16,500 stars in a single week.
Read the story →— Astra for Law pairs GPT-6 Astra with a legal search index of 230M+ URLs and scores 54% on the Vals AI legal research bench — but it still gets nearly half the questions wrong.
Read the story →— OpenAI will retire custom GPTs on December 11, 2026 and replace them with plugins: instructions carry over as a skill, but custom actions and the selected model must be rebuilt.
Read the story →— Faraday Future unveiled nine AI-powered robots across five models at its September 19 event — including the $9,990 Master Mini humanoid and a $129,900 flagship — all on sale now.
Read the story →— Anthropic plans to go public in November at a reported ~$2 trillion valuation that could raise up to $100 billion — while OpenAI forecasts $278 billion in negative cash through 2030.
Read the story →— Anthropic and Accenture will each invest at least $1 billion over five years to put independent evaluators inside Anthropic — testing its frontier models with employee-level access.
Read the story →— The three leading AI labs have been quietly coordinating on AI safety for weeks, OpenAI’s policy chief confirmed — a rare show of unity with antitrust questions attached.
Read the story →— Install it from Microsoft Marketplace and draft, rewrite, and proofread without leaving the document.
Read the story →— A native omnimodal model with a 1-million-token context window, available now in Qwen chat and via an OpenAI-compatible API.
Read the story →— The executive order doesn’t require a kill switch yet — it gives experts until mid-November to recommend how one would work.
Read the story →— OpenAI is testing Sponsored Agents in ChatGPT: labeled brand-run agents that take over after you tap an ad, alongside a ChatGPT Ads Manager plugin that builds campaigns from natural-language prompts.
Read the story →— Google launched an official Google Home MCP server that lets ChatGPT, Claude, and other AI agents control smart-home devices — currently limited to US subscribers on the $20/month Premium Advanced plan.
Read the story →— Anthropic is merging Claude chat and Cowork into a single interface and launching Claude Docs and Claude Slides in beta — rolling out first to Pro and Max subscribers on web, desktop, and mobile.
Read the story →— Salesforce's new Agentforce lineup — Casey, Paige, Carter, Marshall, Piper, Fin, and Hunter — is designed to start work on day one, no assembly required.
Read the story →— Komodor’s new platform tackles the unglamorous half of the AI agent boom: keeping agents governed, grounded, and cost-controlled once they leave the pilot phase.
Read the story →— The three leading AI labs have been quietly coordinating on AI safety for weeks, OpenAI’s policy chief confirmed — a rare show of unity with antitrust questions attached.
Read the story →— The new Model Context Protocol server lets coding agents like Claude and Codex draft, publish, and measure product release notes — with a human approving every post.
Read the story →Three rules, no exceptions:
Categories we track: AI agents, AI tools, platform updates, and automation & business tech.