AI in 15 — September 16, 2026

There's a new model out today that is four hundred times cheaper than the competition. Not four percent. Four hundred times. And it achieves that by being physically incapable of writing a single sentence.

Welcome to AI in 15 for Wednesday, September 16th, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: Google ships two new live voice models and edges into first place on the one benchmark it should have owned years ago.

That four-hundred-times-cheaper model, which its makers say cannot hallucinate. Mathematically.

A pentesting agent got admin on a company's production code in twenty-five minutes, using a key that had been sitting in public since 2023.

Plus Beijing's answer to the pacing essay, Anthropic cuts Claude Code limits and calls it an increase, and somebody wrote a GPU driver with an LLM that nobody will accept.

Marcus, Google put out two models yesterday. Give me the shape of it.

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. Live is the cheap, fast tier — conversation, turn-taking, and visual grounding, meaning you point a camera at something and it talks about what it's seeing in near-real time. Extended Thinking is the heavyweight, built for multi-step tasks that need actual reasoning rather than quick patter.

And the headline number?

Extended Thinking takes number one on the Speech-to-Speech Quality Index at eighty-two point six. OpenAI's GPT-Live-1 Astra is eighty-one point five, Grok Voice Think Fast is eighty-one point three. So a lead, but a narrow one — I'd call it a nose, not a length.

What's the part that actually changes the experience?

Two things, and they're architectural rather than just bigger. First, tool calls run in the background while the conversation keeps going. Second, Extended Thinking reasons and speaks at the same time — it throws out an early verbal cue, "okay, let me pull that up," then narrates progress while the work happens underneath.

Which solves the thing that makes every voice assistant feel like being on hold.

Exactly that. The dead air is the tell. Every voice agent you've used has it, and it's the reason they feel like software rather than a conversation. Google's fix isn't a smarter model, it's a different plumbing arrangement. Also: ninety-seven languages, detected and switched mid-sentence. You can start in English and finish in Afrikaans without touching a setting.

Does that actually work, or is it a demo?

An Afrikaans speaker living in the US said practising with it was, his word, phenomenal. Though the same thread had somebody calling Gemini the only model that sometimes loses context in literally the next message and injects product links nobody asked for. Both can be true.

Other numbers worth having?

Sixty-eight point six percent on agentic voice performance, ninety-seven point seven on audio reasoning — and thirty-five point one percent on a banking-transaction benchmark, which is the humbling one. Real transactional voice work is still hard. On pricing, Google asserts cost-effectiveness and publishes no per-token number, so I'd hold that as a claim rather than a fact.

And it's shipping now?

Developers today via the API and AI Studio, plus Search Live for consumers and Extended Thinking rolling into Gemini Live, Gmail, Workspace. The comment I'd leave you with came from the discussion thread: Google should have won voice on paper years ago. The data, the chips, the money. And it's only now edging ahead, on one index, by one point.

Right, the cheap one. Marcus, a frontier model that can't write. Explain that to me like I haven't had coffee.

A company called TypeSafe AI came out of two years of stealth to announce something they're calling System One Models — after Kahneman's fast, intuitive System 1. Their first model is named Jev, after the economist William Stanley Jevons. And it does not generate text. At all. You hand it unstructured state plus a set of structured questions — yes-no, multiple choice, a score — and it answers all of them in parallel in one query, returning typed values with calibrated probabilities attached.

So it's an opinion machine with no vocabulary.

That's not a bad way to put it. Their own framing is "a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." Seventy to five hundred milliseconds per call against three to three hundred seconds for LLM equivalents. Input at four point two cents per million tokens, output tokens free, against twenty cents to ten dollars for comparable models. They claim speedups up to a hundred and ninety-three times and cost reductions up to four hundred and forty-four times.

And the hallucination claim?

Because the output space is constrained to typed values, they argue hallucinations and type errors are mathematically impossible, not merely rare. Which is honest — it can't invent a fact because it can't form a sentence in which to invent one. The training method is novel too, something they call Reinforcement Learning for Calibrated Decisions, optimising for honest probabilities rather than human preference.

Okay. What's the catch? There's always a catch.

Every benchmark is in-house, and TypeSafe says so themselves, which I'll credit. But note the comparison LLMs were run through TypeSafe's own constraining wrapper, which is slower than normal decoding. Max two hundred and fifty-five choices. And it cannot explain itself — there's no reasoning to read. The sharpest critique in the thread was that a more honest headline would be "trading general-purpose generation for fast typed inference."

Why does this matter beyond the benchmarks?

Because an enormous share of AI spend right now is a frontier model doing something small and dumb inside a loop. Is this spam. Does this output pass. Which of these five matters. Nobody needs Shakespeare for that. If a purpose-built model does it at a hundredth the cost with a confidence number attached, a big slice of current inference spend has a substitute. One commenter noticed the second-order move — at that price you can use it as a general ranking function. Rank a thousand articles, pick the best five. Nobody does that today because it's too slow and too expensive.

And the Jevons name is the joke.

It's a deliberate wink. Jevons observed that making coal cheaper increased coal consumption. Cheaper decisions almost certainly means vastly more decisions, not a smaller bill. The founder is Diogo Almeida, formerly at OpenAI, where he worked on the instruction-following methods that made ChatGPT actually usable.

Security. Twenty-five minutes.

Firm called Strix pointed its autonomous pentesting agent at Baseten, an AI inference company. The agent enumerated subdomains, found an exposed container registry, pulled an image with no authentication, and eventually found a live GitHub personal access token — not in a file, but in the image's build history, where a command had expanded the token variable inline.

Baked in when?

A March 2023 build. Still valid three years later. And it opened admin and push access on Baseten's main product repo, their GitOps repo — the one that drives their production clusters — and their Homebrew tap, plus read-write on private customer repositories.

That's a full supply-chain position.

It is. Credit where it's due on the response: disclosure July 13th, registry made private the next morning, token rotated by half four that afternoon, deletion confirmed thirty minutes later. Baseten's team went on the public thread and thanked them.

So what's the lesson? Rotate your tokens?

That, and a specific technical one: deleting a file doesn't clean your build history. Docker remembers. Use proper secret mounts and set expiry dates. But the real point is the one the top commenter made — this isn't something an unmotivated human would bother to look for, but absolutely something a human could find if they were interested.

So nothing became findable. Looking just got cheap.

That's the whole shift. Every stale artefact anybody has ever published is now economical to search. The vulnerability was three years old and the only thing that changed is the price of attention.

The pacing story. We've run it for four days, so keep me honest and keep it tight. Beijing answered.

They did, and harder than the Foreign Ministry line we covered Monday. State media experts called Amodei's essay inappropriate, hostile and groundless, arguing the real goal is curbing China's AI development through technological barriers. The line that travels is from China Daily: the proposed coordination among three companies sounds rather like a club whose membership rules have been drafted before the guest list is announced. A global AI-safety framework that excludes China is not quite global.

Marcus, that's a good line. Is it a good argument?

It's got a real point inside it — three American companies agreeing something is not a global framework, and Amodei's own step three says so. But weigh the two responses properly. One side has committed to seating independent evaluators inside the lab with badges, desk access and the right to publish. The other side has published an editorial. One of those is checkable.

Fair.

And I'd add the awkward timing, which brings us to the next item.

Anthropic cut Claude Code limits. In the same week.

The temporary fifty percent boost that had been running since May expired on the thirteenth, replaced Monday by what's billed as a permanent twenty-five percent increase. Both framings are literally accurate and describe one number. It is twenty-five percent above the pre-May baseline. It is also a seventeen percent cut versus what you had on Sunday.

And users have noticed.

Cancellation threads trending, some subscribers publicly moving to OpenAI's coding products. The specific pain is that a single long agentic task on the best model can eat eighty to ninety percent of a weekly allowance. So seventeen percent isn't a rounding error — it's the difference between finishing a job and not.

Is there a bigger read here?

This is the unglamorous economics under every agent demo. Long-running autonomous coding is the capability every lab is selling and the most expensive thing a subscriber can do. The question I'd hold is whether "permanent twenty-five percent increase" is a baseline that quietly gets re-anchored again next quarter.

Quick business one. OpenAI bought a camera company.

Glass Imaging, computational photography, more than three hundred million dollars — roughly triple its valuation at the last round. Founded 2019 by two ex-Apple imaging engineers who worked on the technology behind Portrait Mode. Their approach is training networks on the specific optical quirks of an individual camera module, so the picture comes out right at capture rather than being repaired afterwards.

This is the Jony Ive device, isn't it.

Almost certainly adjacent to it. The skeptics had the sharper read though — camera modules are commodity parts with thin margins, so this is likely buying a scarce team rather than a product. There are very few people alive who have shipped flagship smartphone imaging.

And why does a chatbot company need them?

Because a device that's meant to see for you needs an image pipeline good enough that the model isn't reasoning over mush. Buying the camera people before you ship the camera tells you something about what that device actually is.

And one that lands right next to the pacing conversation. Shanghai.

Shanghai AI Laboratory released a seven hundred and forty-four billion parameter mixture-of-experts agentic model called Atria Dawn Preview. No blog post, no press release. The weights just appeared on Hugging Face on the eleventh, three days before the paper. MIT licensed — so any team with enough GPUs gets a deployable frontier-class agentic model with zero licensing negotiation.

Built for what?

Long-horizon agentic work. Problem analysis, tool use, code, running experiments, and explicitly failure recovery. Two things I'd probe. One, the release style — shipping frontier-scale open weights like a routine commit is a completely different competitive posture than a launch event. Two, these are self-reported benchmarks from a state-affiliated lab with no independent evaluation yet. "Competitive across multiple benchmarks" is doing a lot of work in that sentence.

What's checkable?

That the weights are downloadable. That part's real, and it lands the same week Western labs are discussing coordinated speed limits. Which is Amodei's step three, made concrete and uncomfortable.

Last one, and it's a proper fight. Somebody wrote a GPU driver with an LLM.

A developer built a fully OpenGL ES 3.0-conformant Linux GPU driver for the M4 Mac Mini in about a month. Kernel driver, user-space stack, shader compiler. Minecraft at two hundred and twelve frames a second. The kernel driver took three days; the hard part was weeks of firmware reverse engineering. LLMs did the bulk of it, and his honest assessment is that the models were excellent at systematic debugging but got pedantic, chasing completeness over progress.

And the problem is that nobody wants it.

Asahi Linux — the project doing Apple Silicon Linux — has a hard no-generative-AI policy. Their legal cover is clean-room reverse engineering, and their position is that LLMs are likely to taint that. There are contested claims flying around in the comments about the author's history; I'd leave those alone. The solid part is the policy and the achievement.

So the code exists and has nowhere to go.

Reverse engineering undocumented hardware used to cost years of expert time. It now costs weeks. But the project that most needs this driver has ruled it inadmissible, and their reason is legal, not aesthetic. Expect AI-assisted forks that never upstream. Meanwhile Asahi's own M4 support has no GPU driver at all.

One to watch: whether Google DeepMind or Meta say anything about Amodei's coordination proposal. Altman's in, Microsoft's in, and the silence from the other two is now the loudest thing in the room.

Counter: I'd watch the Kalshi story. Semafor reported yesterday that Commerce ordered the prediction market to pull its AI-compute futures curve; Commerce says on record that's flatly false. Whichever way that resolves matters more, because a public price signal for compute is the one thing that would let outsiders check whether this boom is rationally priced.

That's your AI in 15 for today. See you tomorrow.