A house newsletter · daily · researched and written by the Kurzia AI
Wednesday 11 November 2026 · Issue 50 · 5 min read
Amazon drops a frontier model for a classifier (98% cheaper), Gemini 4 Carbon goes GA, Anthropic prices its IPO
What's hot
Strands Agents · yesterday
Strands Decider
Takeaway: about $500 through the decision model
Two engineers on Amazon's internal ticketing system took Claude Sonnet 4.6 off the job of sorting tickets into 20 categories and handed it to Strands Decider, a "decision model" that cannot write prose at all: you give it the ticket and the list of options, and it returns a probability for each plus a confidence score. On their traffic the small model matched the big one's quality, answered much faster and cost a fraction as much. They did not cut over blind — a config flag ran both models on live tickets and every disagreement was judged by hand.
This is the cost conversation your clients are already in, with a shape you can reuse: closed-list classification is the first job to move off a frontier model, and the dual-run switch is the part worth selling. At list prices a million tickets ran about $22,200 through Sonnet and about $500 through the decision model, with the average call falling from 4,215 ms to 897 ms.
“Running both on real traffic, then judging every disagreement, told us more than any offline eval would have.”
New tools & products
Claude Meetings — lets Claude sit in a video call as a labelled participant and start work while people are still talking — drafting the plan being discussed, pulling up last year's figures, answering in the meeting chat. Expect to be asked about it: this is Anthropic pushing from coding tool to office suite, and it lands straight in the meetings your clients' legal teams fret over. Beta on Team and Enterprise plans, Google Meet only at launch, and it leaves on its own if an external guest joins. TechCrunch
Modern Web Guidance — a free set of agent skills from the Chrome, Edge, Firefox and MDN teams that steers coding agents away from dated JavaScript workarounds when the browser now has a native answer. A one-command fix for a complaint you hear on every front-end engagement: agents reach for legacy patterns because that is what their training data is stuffed with. It installs with a single npx command and covers 250 modern web features and 203 real-world use cases, deliberately pruned to stay cheap on context. GitHub (GoogleChrome)
Amp Boxes — always-on small virtual machines you ask Amp's agent to create, then deploy its finished apps onto, each app with its own URL and no containers in between. The cheapest honest answer to "where does the thing my agent built actually live", which is exactly where internal tools usually die. The smallest box is $18 per 30 days, the default one-vCPU, 4 GB box $30, and each one comes with an agent running on it you can ask to fix live data. Amp
ChatGPT Business Agents — verified agents that answer for a brand inside ChatGPT, alongside a Business Hub where a company claims itself and sees how often ChatGPT mentions it. Marketing will bring this to you before engineering does: it is the start of brands buying their way into the answer, and of agents talking to agents on the buying side. In beta for eligible US e-commerce and online service businesses, and OpenAI will not say how it verifies who owns a brand. The Verge
What people are saying
evanxmerz.com · yesterday
Ex-Meta engineer says AI-built codebases quietly turn brittle
Takeaway: at least 90% of his code
Evan X. Merz, laid off by Meta in May, argues that heavy AI use leaves a codebase fragile rather than fast: tests that only cover the easy path, helper layers nobody asked for, and pull requests arriving quicker than anyone can read them properly. His warning signs are operational rather than aesthetic — more incidents, and slower recovery because fewer humans know how the thing works. It set off a long argument on Hacker News.
This is the counter-case a client's engineering lead will raise, and it deserves a straight answer: who still understands this system, and what happens at three in the morning? Merz says he produced at least 90% of his code with AI in his final months at Meta, which he describes as the floor for a decent performance rating.
“AI codebases are software made by AI for AI.”
simonwillison.net · 2 days ago
Agent PR write-ups are so thorough nobody reads them
Simon Willison, whose projects now change mostly through pull requests opened by agents, has started collapsing the agent's write-up behind a click using GitHub's details-and-summary tags, leaving one human-written sentence visible on top. He labels the hidden block with the model and effort level that produced it. AgentConn reads the same week as a broader shift, with engineers tending agent output rather than producing code themselves.
A free habit to drop into a client's review guidelines: the real check is that a human read the change well enough to write one honest line about it.
“My promise to the reader is that the short opener will represent my human judgement as to what is definitely worth their time.”
News
AlphaProof Code — a Google DeepMind framework in which agents write machine-checked mathematical proofs that real code holds its security promises — used on Signal's encryption code, finished in four weeks, with humans reviewing about one percent of it. It is the strongest answer yet for a client asking how you would ever prove agent-written code is safe, and the artifacts are being open-sourced. The agents emitted over six million lines of Lean proof and ran on Gemini 4 Carbon, which Google made generally available on 10 November at $2 per million input and $10 per million output tokens. Google DeepMind
Anthropic IPO — The roadshow is live: 825 million Class A shares at an indicative $70 to $78, listing on Nasdaq under the ticker ANT. Clients will read the debut as a verdict on whether all this agent spending pays, so know the numbers before the meeting. That reaches as high as a $1.4 trillion valuation according to the New York Times, well under the $2 trillion once floated, and the Wall Street Journal reports shares are meant to start trading Thursday or Friday next week. Anthropic
Curative — a US health insurer running an agent called Gwen that works a provider deal end to end and signs around 2,400 contracts a week, more than its whole 2025 total of 2,100. The case study to show a client who thinks agents only write code: this is a revenue process, and the human team is still paid on the agent's deals. Anthropic / Claude
Fireworks AI — The model-hosting company now says the intrusion it owned up to last month went further than first reported, touching billing and account records, secrets, datasets and unfinished models on a limited number of customer accounts. If a client's fine-tuning or weights sit with a third-party host, this is the supplier question of the week. Its first notice went up on 10 October 2026, reporting unauthorised use of internal systems. The Hacker News
State of Developer Ecosystem 2026 — JetBrains' tenth annual survey asked nearly 29,000 developers in 198 countries how they work; for most employed respondents, agents are involved in the majority of their tasks, past the 60% mark. Useful ammunition when a client still treats agents as a pilot — and the ROI figures show nobody has the measurement side worked out. 51% of tech managers track token usage, but only 30% call it one of their most insightful metrics. JetBrains
One takeaway
Quote cost per task and evidence, not model names. Three of this week's items point the same way: the frontier model is no longer the default answer. One team moved a fixed-choice job onto a tiny classifier and kept the quality; another had agents produce proofs that real code holds; a third keeps finding that the slow part is reading what agents hand back. If you advise on agentic delivery, the thing to put in the proposal is the cheapest component that does the job, plus the evidence it did it. Strands Agents