Open models, voice and images come next. Xiaomi's MiMo-V2.6-Pro is the smartest open model right now, and it has a trillion parameters. ElevenLabs and Google released new voice models, and Google's Nano Banana 2.1 puts it back near the top for images.

The agents chapter covers OpenAI's Dots and Space in ChatGPT, Meta's Muse, xAI's Grok Bot, Microsoft's Autopilot, Codex usage resets and Anthropic's API credits for subscribers. The industry chapter looks at the a16z list of the top 100 consumer AI apps. I skip most of the security news and the debate about mathematics results.

Key Stories

• Opus 5.5, Sonnet 5.5 and Haiku 5.5 give Anthropic a full range again, and Opus 5.5 is cheaper than the previous Opus and above Fable on the Artificial Analysis Intelligence Index.
• GPT-6.1 Sol arrived a week after GPT-6 Sol and Luna and is much closer to Astra on the benchmarks.
• Muse Wonder scores higher than Astra on the composite intelligence index, with good early feedback on coding.
• Jev returns scores for decisions instead of generating text, and answers around 200 times faster than a large language model.
• MiMo-V2.6-Pro is Xiaomi's trillion-parameter open model, released with a public view of its training run.
• Dots are assistants inside ChatGPT, available from $100 a month, that connect to apps, read email and talk to you on Teams and Slack.
• a16z's Top 100 Consumer AI Apps includes a ranking by monthly revenue from US consumer card spending, where OpenAI and Anthropic lead.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

The practical examples include running smaller models locally, voice conversations that delegate work to other models, and agents that carry out scheduled tasks or communicate across sessions. I also look at transcription, image generation and world models.

The later chapters cover NVIDIA's agreement to acquire Hugging Face, Stripe's agreement to acquire OpenRouter, and the security incidents reported during model evaluations. Those incidents involved different failures in access and containment. They also prompted discussion about independent evaluators' access to the labs. The roundup closes with the debate about AI-generated mathematics and how results should be checked and explained.

Key Stories

• Claude Fable and Mythos 5.1 and GPT-6 Astra improve science and computer-use results, while raising questions about benchmarks, safeguards and access.
• Muse Glimmer and Qwen3.8-27B add options for running models on personal hardware, with memory requirements that depend on the configuration.
• GPT-Live-1 lets a voice model delegate deeper reasoning and tool calls while handling the conversation.
• OpenClaw 2.0 simplifies setup, while Muse offers a personal agent with its own computer.
• METR's Hugging Face investigation examines how agents in an evaluation communicated and acted beyond their assigned tasks.
• A Severe Misalignment of AI in Mathematics argues that solving a problem must be followed by the work of explaining its methods and developing understanding.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

Microsoft Build put workplace context, governed agents and Copilot into one enterprise frame. OpenAI's new ChatGPT memory system was the same problem from the consumer side. Claude Code, Codex, Antigravity and Cursor showed the practical tool shift: instead of asking a chatbot for an answer, you are setting up work for agents.

The model section was more mixed. Microsoft launched its own MAI model family. MiniMax M3 pushed the open-weight frontier. Gemma 4 12B and Google Eloquent showed why local multimodal and voice workflows matter. The later sections covered pricing and compute, AI in research and security, and public arguments about writing, detection and institutional overreach.

Key Stories

• Microsoft Build and workplace context — Microsoft IQ, Work IQ, Fabric IQ and Scout put enterprise agents, context and governance into one stack.
• ChatGPT memory and dreaming — OpenAI rebuilt memory around background context synthesis, which makes continuity a product feature rather than a saved-notes feature.
• Claude Code dynamic workflows and Codex for knowledge work — coding agents are becoming workflow tools for research, documents, spreadsheets, contracts and other knowledge work.
• Google Antigravity 2.0 and Google AI Studio mobile — Google is converging on the same agent-first tool shape as Codex and Claude Code.
• Gemma 4 12B and Google Eloquent — local multimodal AI is becoming practical enough to show up in voice, dictation and editing workflows.
• GitHub Copilot consumption pricing and Claude Agent SDK accounting — better agents use more compute, and flat subscriptions are getting harder to sustain.
• Project Glasswing and Mythos, OpenAI Rosalind and Google AI Co-Scientist — AI is moving into security, biodefense and research workflows.
• Pangram and AI detection — the writing and detection debate needs institutional modesty, especially when tools are used to judge individual students or individual sentences.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

Voice and image became more usable. Gemini 3.1 Flash TTS introduced audio tags — bracketed natural-language directions like `[excitement]` embedded in the script — and is free in Google AI Studio. GPT-Realtime-2 brought full flagship intelligence into the voice stack, with passive listening, parallel tool calls, and live translation. GPT-Image-2 took the #1 slot across the Image Arena leaderboards by 242 ELO points, ending the brief period in which Google's Nano Banana family had led image generation.

Agent platforms appeared across several labs at once. Anthropic's Managed Agents added memory and a multi-agent shared-memory feature it calls "dreaming." OpenAI previewed Workspace Agents. Microsoft launched Foundry Hosted Agents. Mistral shipped Workflows. Cursor opened a Cloud Agents API. Codex itself was upgraded into something closer to a general work environment with browser previews, sidebar artifact rendering, and image generation. The infrastructure items were Anthropic's compute partnership with SpaceX, Microsoft and OpenAI's amended non-exclusive partnership, and OpenAI models becoming available on Amazon Bedrock for the first time. The labour section covers Alex Imas, Dylan Matthews, Anthony Pompliano, and Ethan Mollick on jobs, risk, and professions.

Key Stories

• Claude Opus 4.7 and GPT-5.5 — the two top-tier model releases, with different reception and different strengths.
• GPT-Rosalind — OpenAI's life-sciences model, now looking for academic testing partners.
• Qwen 3.6, Kimi K2.6, and DeepSeek V4 — the open-weights tier keeps shipping at speed.
• Gemini 3.1 Flash TTS and GPT-Realtime-2 — voice models get cheaper, smarter, and less awkward.
• GPT-Image-2 and Claude Design — image generation and code-as-design become more useful for real design work.
• Codex for almost everything — Codex moves toward a general work environment with browser previews, sidebar artifacts, and image generation.
• Anthropic Managed Agents, OpenAI Workspace Agents, and Microsoft Foundry Hosted Agents — agent platforms become standard lab infrastructure.
• Microsoft Copilot Agent Mode and Claude across Microsoft 365 — Microsoft 365 added more agent and model-choice features.
• Mythos Preview / Project Glasswing and Where the goblins came from — safety research becomes more concrete, from cyber capability to model personality drift.
• Anthropic-SpaceX compute partnership and Microsoft-OpenAI partnership amended — compute and cloud access remain the hard industrial constraint.
• Alex Imas, Anthony Pompliano, Dylan Matthews, and Ethan Mollick — jobs, AI risk, and professions.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

The biggest announcement was Anthropic's Claude Mythos, revealed under Project Glasswing — a model so capable at chaining security vulnerabilities that Anthropic chose not to release it publicly. Meanwhile, real-world supply chain attacks hit axios (300M weekly downloads) and litellm in the same week, making the security conversation very concrete. OpenAI's acquisition of Astral (Ruff, uv) raised questions about who controls critical open-source infrastructure.

Coding agents became the new normal. Codex crossed three million weekly users. OpenAI matched Anthropic's pricing ladder with a new $100/month tier. Cursor redesigned around autonomous agents. The OpenClaw ecosystem matured fast enough that Microsoft hired a VP whose entire job is OpenClaw strategy. And Karpathy's AutoResearch showed what happens when you let agents run research in loops for hours — with results ranging from 53% performance gains to 103 experiments yielding one improvement.

Key Stories

• GPT-5.4 Mini and Nano — OpenAI's mid-tier model catches up with Gemini Flash, now with proper agentic capabilities.
• 1M context GA — Anthropic makes the full context window available at standard pricing for Opus and Sonnet 4.6.
• Gemma 4 — Google's open model family, Apache 2 licensed, multimodal with audio, runs on consumer hardware.
• Project Glasswing / Claude Mythos — Anthropic's most capable model, restricted to security partners for vulnerability discovery.
• GLM-5.1 — first open-weight model to top SWE-Bench Pro, designed for long-horizon agentic work.
• Gemini 3.1 Flash Live — smart, cheap live voice model with video — a step change for voice AI.
• Supply chain attacks on axios and litellm — social engineering and credential theft hit packages with hundreds of millions of downloads.
• AutoResearch — Karpathy's autonomous research loops, with community results across 6 domains in 2 weeks.
• Codex hits 3M users; $100/month tier — the new pricing normal for coding agents.
• OpenClaw memory-wiki — the agent memory problem gets its most principled solution yet.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

The coding agent arms race accelerated further. Anthropic raised $30 billion in Series G funding, driven largely by Claude Code adoption. Cursor's agents can now test their own work by controlling a virtual machine. Claude Code got voice mode, security scanning, Figma integration, and auto memory. Google released a Workspace CLI designed not for humans but for agents to use directly.

On the agent side, METR measured Claude Opus 4.6 reliably handling tasks that take humans 14+ hours. Notion launched custom agents, bringing autonomous workflows to a mainstream productivity tool. Perplexity Computer orchestrates 19 models in parallel for complex autonomous tasks. And practitioner writing on agent design, harness engineering, and AGENTS.md instruction limits started shaping a new discipline around working with agents rather than just using them.

Key Stories

• GPT-5.4 — the best model available by many accounts, with steerability, better tool use, and 83% on GDPval (up from 71%).
• Gemini 3.1 Pro — Google's strongest model for one-shot generation of visual and graphical artifacts.
• Claude Sonnet 4.6 — Anthropic's middle-tier model with significantly improved computer use skills.
• Anthropic raises $30 billion — growth driven by coding agents, not chatbot popularity.
• METR: Opus 4.6 handles 14.5-hour tasks — the task complexity ceiling for autonomous AI work keeps rising fast.
• Perplexity Computer — cloud-based autonomous agent orchestrating 19 specialized models.
• Agents can follow ~150 instructions — practical limits on AGENTS.md files, with implications for prompt design.
• OpenAI abandons SWE-bench Verified — the standard coding benchmark retired because it can't keep up.
• Anthropic refused Pentagon demand to remove safeguards — while OpenAI signed a classified DoD agreement.
• Ads coming to ChatGPT — OpenAI testing ads in the free tier for US users.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

On February 5th, Anthropic and OpenAI released new models within minutes of each other — and both were explicitly positioned as going beyond coding. OpenAI even ran a Super Bowl ad for their Codex app instead of promoting ChatGPT.

Meanwhile, personal AI agents hit the news with OpenClaw and NanoClaw, and an AI-only social network called Moltbook became one of the most talked-about experiments.

As always, there were many other new releases of new tools and models. OpenAI released Prism for scientists writing papers in LaTeX and an agent tool for enterprises called Frontier.

Key Stories

• Claude Opus 4.6 and GPT-5.3 Codex launched on the same day — both moving beyond coding into general computer work. Opus 4.6 | GPT-5.3 Codex
• OpenAI's Super Bowl ad was for Codex, not ChatGPT — a $10-20 million bet on AI-as-builder.
• An Anthropic researcher built a C compiler with 16 AI agents — 100,000 lines of code, $20K, no human coding.
• Moltbook: an AI-only social network — autonomous agents founding communities and debating philosophy. Moltbook | Scott Alexander's analysis
• Mitchell Hashimoto went from AI skeptic to agent-first — "Immediately cease trying to perform meaningful work via a chatbot."
• 4% of GitHub commits now come from Claude Code — projected 20%+ by end of 2026.
• Apple partnered with Google to use Gemini as Siri's backend.
• Kimi K2.5 from Moonshot AI — open-source, 1 trillion parameters, "agent swarms" built in.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.

📅 Next session: MondAI Roundup — November 2026 — Monday 16 November 2026 at 12:30 GMT. Register here.