Two months in one dispatch, and a lot happened. GPT-6 and Claude 5.5 landed, more training agents got caught hacking, and even Dario’s saying the industry needs to slow down. Meanwhile, Nvidia bought Hugging Face, Stripe bought OpenRouter, and Jev introduced Decision Models.

AI Models & Releases

  • GPT-6 Astra: OpenAI’s new flagship, billed as both the smartest and the most aligned model out there, with near-perfect scores on ARC-AGI-3 and FrontierMath Tier 4. They also claim it went out of scope 0% of the time on a new eval built after the Hugging Face incident (keep that number in mind for the next section).

  • GPT-6 Sol and Luna: The cheaper tiers of the GPT-6 family, trained the same way as Astra and priced 50% lower than their GPT-5.6 versions.

  • GPT-6.1 Sol: A week later, Sol got an upgrade that nearly matches Astra on coding, computer use, and professional work at a fifth of the price. I didn’t have time to test the 6 but had a lot of fun and great results with 6.1 generating 3D simulations in WebGL.

  • Claude Opus 5.5: Anthropic’s first model since calling for a pace on the frontier. It performs around Fable 5.1 level on most work, costs about 40% less to run than Opus 5, and scores best on their alignment audit so far. And it finally is less verbose and annoying to work with!

  • Claude Sonnet 5.5: The faster, cheaper sibling. Same price as Sonnet 5 but 30%+ faster and up to 30% cheaper per task, and it’s the first Sonnet to ship with cyber safeguards. Looking forward to see if they can get this sort of upgrade to Haiku soon.

  • Jev and System One models: TypeSafe AI built a new kind of model that takes text in and only returns typed decisions with probabilities (yes/no, choices, scores). Output is free and input costs $0.042 per million tokens.

  • Atlas: A World Model for Spatial Intelligence: World Labs’ new world model works natively with text, images, video, and 3D, and can generate up to a minute of 1440p video with precise camera control. Very impressive examples.

Agents Behaving Badly & Pacing the Frontier

  • OpenAI’s rogue agents: More fallout from OpenAI’s training runs. Agents doing a web research benchmark turned public wikis into a private message board, and another report traced 2,000+ malicious RubyGems packages back to OpenAI agents trying to steal API keys. The UK’s AISI then found GPT-6 Astra completed unsanctioned supply-chain attacks in 29.2% of simulated cyber evals, and Eoin Higgins argues there are no “rogue” agents, only companies that didn’t restrict them.

  • Why are AI agents lying, cheating and coordinating?: Yoshua Bengio’s take on why all this keeps happening. Training rewards whatever gets the task done, so patching each bad behavior risks selecting for models that cheat without getting caught.

  • We Must Pace the Frontier: Dario Amodei now says the industry needs to slow down how fast capabilities improve. His two reasons are AI starting to build the next generation of AI, and the OpenAI and Hugging Face incident. The industry has mostly rallied behind him since.

  • Our framework for reporting model misalignment: OpenAI commits to publishing misalignment incidents quickly, even before they understand them, starting with six reports. They also say the industry hasn’t solved alignment well enough to keep scaling at max speed for much longer, which is quite the admission.

  • AI used to design brand new viruses: Stanford researchers used AI to design complete virus genomes from scratch, and 16 of them actually worked at killing E. coli. Great news for treating antibiotic-resistant infections, and a pretty obvious biosecurity worry.

Tech & Industry

  • NVIDIA to Acquire Hugging Face: After talks leaked in late August, Nvidia is buying Hugging Face for about $12.9B. Jensen promises it stays open and won’t require Nvidia hardware, but its neutrality was the whole point of Hugging Face. I hope they won’t ruin it.

  • OpenRouter is Joining Stripe: Stripe is buying the model gateway, reportedly for $7B+. OpenRouter says nothing changes for users and routing stays neutral. We’ll see.

  • AMD acquires Taalas: The Toronto startup etches model weights straight into silicon, and its first test chip served Llama 3.1 8B at nearly 17,000 tokens a second. I mentioned Taalas back in February, so it’s cool to see them land somewhere big.

  • OpenAI cuts off Cursor after the SpaceX acquisition: OpenAI is ending its models in Cursor on November 12, saying it “cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts.” Ouch.

  • Tailwind Labs is joining Shopify: Tailwind gets a stable long-term home and stays MIT-licensed.

  • Mistral raises €3B: A Samsung-led Series D at a €21B+ valuation, the biggest equity round ever for a European tech company. The pitch is sovereign, open-weight AI with no vendor lock-in.

AI Tools & Agents

  • HarnessTax: How Much Does the Harness Matter for Coding Agents?: Arena tested 21 model and harness combos and found the harness barely changes success rates but can change the cost by up to 5x. Worth a read if you’re looking for the most cost-effective combo of an AI coding harness and model.

  • Loop engineering: Getting started with loops: The Claude Code team defines a loop as an agent repeating cycles of work until a stop condition is met. They walk through turn-based, goal-based, time-based, and proactive loops, plus when each one makes sense.

  • Munder Difflin: An open-source harness that runs Claude Code, Codex, Gemini CLI, and friends as clones in an office, with Michael running the show and handing out tickets. Yes, it’s exactly the Dunder Mifflin bit you’d hope for. (Spoiler alert: I’m a big fan of The Office.)

  • The New MCP Roadmap: The July spec dropped sessions and the initialization handshake so servers can scale without holding state. The new roadmap has five priority areas, including messaging primitives for long-running agent work and agent identity.

  • Previewing the Model Hardware Standard: Anthropic and HHMI Janelia built a standard driver that lets AI agents operate lab and factory hardware like microscopes and robotic arms. It’s model-agnostic, works over MCP, and is meant to go open source after the research preview. I hope one of our customers let’s me test drive this with one of their robots eventually.

  • Mistral x Mozilla: Firefox’s AI assistant, Smart Window (beta), now runs on Mistral models in France and North America. Nice to see two open-source folks team up on private AI browsing.

Security & Privacy

  • Claude’s text watermarks: To comply with the EU AI Act, new Claude models now watermark everything they generate, worldwide, including in Claude Code (The Independent covers the fear of human writing getting flagged). It works by nudging word choices, which John Gruber calls a perversion of writing, since picking “grey” over “overcast” is exactly what writing is.

  • ZCode uploads your git history: Z.ai’s coding app was caught silently uploading the whole workspace, full .git history included, encrypted with a key only Z.ai holds. Iiish, that doesn’t sound good, and it’s a good reminder that open GLM weights don’t make Z.ai’s harness open too.

  • ChatGPT now knows what you do on other websites: OpenAI’s ad pixel sends a cookie tied to your ChatGPT account from every advertiser site you visit, along with what you browse and buy. Facebook ads targeting all over again?

  • Google is buying tons of Spirit employee data: Google won the bankrupt airline’s auction for its employment and workplace records. The flight attendants’ union says the privacy protections cover customers, not the workers whose emails are in there.

Odds & Ends

  • WeatherNext cyclone forecasting: DeepMind’s model gives forecasters roughly an extra day of warning on cyclones, about a decade’s worth of progress. They’re open sourcing it too.

  • Stalking the Wily Hacker: 40 years later: Cliff Stoll, of The Cuckoo’s Egg fame, back on stage 40 years later. The HN comments are worth it on their own, full of people sharing stories about buying Klein bottles from him.

  • fable51-worlds: Claude Fable 5.1 agent swarms take a sentence, photo, or film clip and build a walkable Three.js world from it (Union Square, Kyoto, even the Death Star trench run), all in code with no game engine. Gives me Sims vibes but I love it.