The Signal — September 21, 2026

Three items today about what happens in the gaps: the gap between a model's answer and the code that has to act on it, the gap between an incident and the disclosure of it, and the gap between a voice assistant talking and a voice assistant working.

The fastest-adopted model on Vercel's gateway does not write anything

A week ago TypeSafe AI released Jev, and Vercel now reports that it has been picked up faster than any model in the history of its AI Gateway. Within twenty-four hours it was running in almost 13 percent of paid teams, twice the GPT-5.6 family and more than six times Fable 5.1. Cloudflare's Workers AI carries it too. The company was founded by Diogo Almeida, previously a researcher at OpenAI.

What makes the adoption curve interesting is that Jev does not generate prose. An application hands it a piece of state, say a support ticket, along with a map of typed questions: is this urgent, which team should handle it, how frustrated is this person on a three-point scale. Jev evaluates all of them in one parallel pass and returns typed answers with probabilities attached, in a shape the surrounding code can branch on directly. No tokens to parse, no JSON to repair, no retry when the model decides to explain itself first.

TypeSafe is betting that agent budgets are mostly spent on decisions, not prose. Most of what an agent asks a frontier model to do is not writing, it is choosing: which tool to call next, whether to retry, whether a command is safe to run, whether this case needs a human. Those decisions have been getting made by models built to compose paragraphs, at paragraph prices and paragraph latency. TypeSafe says its own workflow evaluations put Jev up to 193.6 times faster and 444.6 times cheaper than language models on the same calls. Treat that number the way you would treat any vendor's benchmark of its own product against a category it is trying to displace. Vercel's changelog is refreshingly blunt about the limit, telling developers to calibrate the returned probabilities against their own labeled examples before trusting them. The adoption figure is first-party too, but it is Vercel counting its own paying customers, which is a different kind of claim from a benchmark score.

Sources: Vercel · Vercel Changelog · Cloudflare · Techmeme

Google knew for seven weeks and decided it was not news

In May, during a capture-the-flag security exercise run by the Israeli evaluation firm Irregular, a Gemini model got into the systems of three real companies. It guessed a password until one let it in, and in the other two it found credentials sitting in public repositories. Each time, once it worked out the target was a real business rather than part of the exercise, it stopped. Irregular has since attributed the whole class of incident to a naming mistake, a fictional company used in the exercises that happened to match a live domain, combined with internet access the models were never supposed to have.

Irregular told Google at the end of July. Google told the three companies and told federal authorities. Then, for seven weeks, it told the public nothing, until the Wall Street Journal came asking on September 18. Asked to explain the silence, Google said the model had behaved appropriately, that it had stopped on its own, and that the episode was not an example of misalignment and did not warrant disclosure.

The break-in is the less interesting half. OpenAI, Anthropic and Meta all had incidents with the same evaluation firm this year, and all three disclosed before a reporter forced the question. Google is the first to be asked first. Its reasoning is internally coherent, which is what makes it worth writing down: a company can define misalignment narrowly enough that a model autonomously breaching three companies it was not authorized to touch falls outside the definition, because the model stopped when it noticed. Every lab currently discloses these events voluntarily, on its own judgment of what counts. Google has now shown what the low end of that judgment looks like, and there is no rule that says it has to be higher. Google has published no document of its own about any of this. Everything attributed to it comes from statements handed to reporters.

Sources: The Verge · Reuters · The Guardian · Al Jazeera

Tencent split the talking from the thinking

Voice assistants take turns. You say something, it responds, you go again, and if you ask it to do anything substantial the conversation stops while it works. Researchers from Tencent's Hunyuan Speech team and several universities published a system called Gander that tries to get out of that pattern by running two models instead of one. A small fast component they call the cerebellum handles the conversation, deciding every second whether to listen, speak, or stop speaking, holding about two minutes of context. Behind it, a swappable brain does the slow work: searching files, writing code, fixing a bug. You can keep talking to the front half while the back half grinds, interrupt it, or change the task partway through.

The brain is a socket rather than a model. Codex or Claude Code can be dropped in without retraining the conversational half, and in the published tests an unspecified model from the GPT-5.6 family filled the role. That is the part with legs beyond this particular paper: it makes conversational timing a separate engineering problem from task competence, so improvements to either one land independently.

On Full-Duplex-Bench v3, across all 100 scenarios, Gander started speaking at the right moment every time and cut the user off in 8 percent of cases, against 13.5 percent for GPT-Realtime and close to 48 percent for the worst system in the comparison. Its task accuracy came in just below the weakest competitor. The authors are unusually honest about this, noting that the benchmark scores the whole pipeline including speech recognition and audio output, and that the brain does considerably better when handed text directly. They also flag that Gander performs worse than its own base model on video and audio understanding, which they attribute to training that favored conversational fluency over precise perception. All of these numbers are the authors' own, on a benchmark of a hundred scenarios, and the model is not out yet. Weights and training data are promised; the code repository is already up.

Sources: arXiv · The Decoder · GitHub

On the Editor's Desk

A few things we looked at and left out. Epoch AI and Ipsos published good polling showing near-daily AI use among US adults jumping from 8 to 19 percent since March, but we used Epoch's data three days ago and did not want to go back so soon. Microsoft's StudentSim work on simulated learners is the strongest research in today's pool, and we have circled it twice already this month; the paper itself went up on September 1, so the article that prompted us was three weeks behind it. Trump's remarks about creating an "AI Force" and appointing an AI czar are confirmed by several outlets, but there is no order, no mandate and no named appointee, so there is nothing yet to report beyond the sentence itself. Alibaba's Qwen-Image-2.1 was the last cut, purely for space: a 7-billion-parameter open-weight image model that runs on a consumer graphics card, released under a license that forbids commercial use. It may show up tomorrow.