- Sakana AI shipped Fugu Max and Fugu Ultra v2, a single endpoint that farms your request out to a pool of other labs' models. Sakana says the top tier hits its scores without the two biggest frontier models in that pool.
- Google gave Windows a native Gemini app, summoned with Alt plus Space, free with daily caps on the base model.
- California created the country's first AI auditing framework: a state registry of auditors, plus a ban on selling an AI audit if you're not on it.
- Anthropic published how it measures whether its own models help with military targeting and conventional weapons work. Useful, and also homework the student graded.
- One thread runs through all of it: the model keeps getting demoted to plumbing. If you'd rather build than watch, that shift is the whole argument in going from generative to functional AI.
Three things happened Friday and none was a smarter model. A Japanese lab sold a router. Google sold a keyboard shortcut. California started licensing the people who'll inspect both.
Put them together and the same sentence falls out. The model is sliding into the position the database holds at a normal company: important, expensive, and almost never the thing you sell. What sits above it, the routing and the shortcut and the paperwork, is where this week's money and risk landed.
The Front Page: Sakana Sold a Switchboard and Called It a Model
Sakana AI, the Tokyo lab, released Fugu Max and Fugu Ultra v2 on September 11. Sakana's own announcement describes both as a multi-agent system delivered as one model: you send a request to a single endpoint, and behind it Fugu picks which models from a pool should do the work, coordinates them, and hands back one answer. Ultra v2 takes text, images, and PDFs, carries a one-million-token context window, and lists at $5 per million input tokens and $30 per million output. Fugu Max is the cheaper tier. A second read covers both.
Here's the claim that makes it interesting. Sakana says Ultra v2 reaches its numbers without Fable 5, Fable 5.1, or GPT-6-Astra in its pool. Translated: the orchestrator reckons it can beat the class without the strongest students enrolled.
Treat that with the usual distance. The scores are Sakana's, on Sakana's tasks, and an orchestrator has an obvious incentive to show it doesn't need the models it would otherwise pay a premium for. There's a plainer business question too: a router's margin is the spread between what it charges you and what it pays the labs underneath, and any of those labs can reprice on a Tuesday.
What it means: If you build products, this is the week the "which model do we standardize on" meeting got less important. Routing across models is becoming a product category rather than a hack. If you buy products, ask a harder question back: when the answer came from a pool, which member wrote it? You'll want that on record before anybody asks you for it.
Releases & Features
Gemini landed on Windows as a real app. Google shipped a native Gemini client for Windows 10 and 11 this week, after the Mac version. Alt plus Space pulls it over whatever you're working on, the core app is free with daily limits on the base model, and Google's pitch is that it stays "lightweight and quiet." Coverage from 9to5Google and BetaNews.
A 2.5-billion-parameter model that fits on your laptop. The small-model lane was quiet this week, so the freshest thing in it is still OpenBMB's MiniCPM5-2B, top of the open-weights models under 4 billion parameters on the Artificial Analysis Intelligence Index at 15, against 11 for IBM's larger Granite 4.2 3B. Weights are Apache 2.0, the training data shipped with them, and it loads in llama.cpp, Ollama, and LM Studio unmodified. Fuller breakdown.
What it means: Google isn't competing on intelligence here, it's competing on muscle memory. A keyboard shortcut beats a benchmark point once people stop thinking about which tab to open. And a model small enough for a laptop, licensed loosely enough to ship inside a product, is the option most teams forget they have. Most jobs need a cheap model that never leaves the building.
In the Lab
Anthropic's Frontier Red Team published work on September 10 describing how it measures tactical intelligence targeting and conventional weapons capabilities in its own models. That team's remit is the uncomfortable end of the field: what a frontier model could contribute to cybersecurity attacks, biosecurity risks, and autonomous systems, tested deliberately rather than found by accident.
The honest framing: this is a lab writing its own exam, sitting it, and publishing the grade. That beats silence by a wide margin. But "we evaluated ourselves and found the risk acceptable" wouldn't survive five minutes in any other regulated industry, which is the thread running into the oversight story below.
What it means: Dangerous-capability testing is becoming a published artifact rather than an internal memo, and that norm is worth defending. For anyone building on these models the transferable lesson is smaller: decide in advance what your system must never do, then test for it on purpose. That's an ordinary step in any serious agent workflow, and far cheaper before launch than after.
The Oversight Desk
California built a profession this week. Governor Newsom signed AB 1405, creating a state AI Auditor Registry and making it unlawful for anyone unregistered to offer, sell, or conduct an AI audit, plus SB 813, which sets requirements for the third-party organizations that certify AI compliance and safety. The Transparency Coalition calls the pair the nation's first AI auditing framework. He also signed AB 1883, barring employers from using AI surveillance tools that collect neural data or infer a worker's emotional state.
Follow the money and the shape gets clearer. A registry creates a licensed trade, and licensed trades bill by the hour. Note what the laws don't do: they don't say who must be audited. They build the inspectors first and leave who gets inspected to other statutes, which is a legislature buying itself optionality.
What it means: Compliance is turning into something you can purchase, which means it's turning into something you can be asked to show. Newsom has until September 30 to act on thirty-odd remaining AI bills, so this is a floor, not a ceiling. If you run AI in hiring, lending, or healthcare, the useful move this quarter is boring: keep a record of what your system decided and why, before somebody with a registry number asks to see it.
Every story today points at the same unglamorous asset: a written record of what runs, in what order, and who signs off. Tell BYOBot about a repeating job and get that map back.
On the Radar
Smaller moves worth a glance, sources attached.
- Anthropic graded four of its own incidents. A September 9 alignment assessment examines four cases in which Claude models gained unauthorized access to real third-party systems. Read it if you run agents with credentials. The assessment.
- Money moved to the pipes. Ayar Labs raised a $150 million Series E extension for silicon photonics, the optical links that shuttle data between chips. Unfashionable, and the reason any of this runs. Roundup.
- Voice agents found a market in Japan. nocall.ai closed roughly 800 million yen in Series A funding from JAFCO, Archetype Ventures, and Coreline Ventures. Details.
- Six states are still writing AI law. Michigan, Pennsylvania, Massachusetts, Ohio, New Jersey, and North Carolina remain in session, on top of 85 AI laws already enacted across 27 states this year. Weekly tracker.
- Senior care drew a $550 million valuation. Inspiren raised a $70 million Series C led by NewView Capital for AI monitoring in senior-living facilities, one of the few places a false negative is measured in ambulances. Funding news.
The Bottom Line
Friday's through-line is demotion. Sakana bets no single model is worth standardizing on, Google bets a shortcut beats a score, California bets the industry will need somebody licensed to check its work. All three treat the model as a component you swap. Watch whether anyone independent reproduces Sakana's claim about that missing frontier tier, and whether Newsom's remaining bills turn the registry into an obligation rather than a directory. The least glamorous asset on this page is also the one you can build today: a plain written record of what your automated work does and who answers for it.
Frequently Asked Questions
-
An orchestrator sits between you and several AI models. You send one request to one endpoint, and it decides which model or combination of models should handle it, then returns a single answer. Sakana's Fugu family works this way. The appeal is that you stop betting your product on one lab. The tradeoff is that you can no longer see exactly which model wrote your output, which matters if you ever need to explain a result to an auditor or a customer. The same coordination problem shows up inside a single product, and it's what we walked through in designing a multi-tasking agent.
-
An AI audit is an independent review of whether an AI system does what its operator claims and meets whatever rules apply to it. California's AB 1405 creates a state registry of AI auditors and makes it unlawful for anyone not on that registry to offer, sell, or conduct an AI audit, while SB 813 sets requirements for third-party organizations that certify AI compliance and safety. Nobody is audited by default yet. The laws build the profession first, and obligations for specific sectors follow from other statutes. If you want a sense of what a reviewable paper trail looks like in practice, our legal ops automation playbook covers intake, routing, and tracking.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
