Andon Labs gave Google's newest model a simulated vending machine business to run, and the model announced that its priority was to, quote, obliterate my competition. It did win the simulation. It's Thursday, October 29th, and this is AI Daily Briefing. On today's list: Google's Gemini 4 Carbon takes the top spot on the main independent leaderboard but goes only to security teams first, Artificial Analysis rebuilds its index and the open-weights gap jumps back to twelve points, Microsoft and Alphabet post quarters that show who is actually collecting AI revenue, and New York signs a kids' chatbot law whose penalties can reach all the way into the model. If today's show was useful, follow us wherever you're listening — the next one will be waiting.
Thomas Claburn, writing in The Register:
Priced initially at $2 per million input tokens and $10 per million output tokens, Gemini 4 Carbon sells for the same amount as its predecessor Argon during an indeterminate introductory period. When that period expires, Google will lift prices to $4 per million input tokens and $20 per million output tokens.
Google DeepMind announced Carbon on Wednesday, two weeks after Argon's wider rollout, which we covered on October 14th. Koray Kavukcuoglu calls it Google's strongest model yet, especially for cybersecurity and long-running tasks. Claburn notes that the post-introductory price matches Claude Opus 5.5, which tells you who Google thinks the competition is. Independent numbers first. Artificial Analysis puts Carbon at fifty-nine on its new index, one point over Opus 5.5 and two over GPT-6 Astra. Argon scored fifty. On the Coding Agent Index, Carbon is second, three points behind Claude Code running Opus 5.5. And Vals AI ranks it second on the Vals Index, behind Claude Sonnet 5.5. Then Google's own table. Sixty-five point two percent on Terminal-Bench 4.0, against just under sixty for both Opus and Astra. Ninety-six point two on the GraphWalks long-context test, where Astra gets seventy-one point eight. It loses FrontierSWE 2 to Astra and OSWorld-2.1 to Opus. The cyber numbers are the real pitch. Eighty-eight point two on SEC-bench Pro for finding vulnerabilities, fifty-four point eight on ExploitGym for exploiting them, seventy-one on CWE-bench for patching. Google says those external runs had no guardrails. That's why the first users this week are a subset of cyber defenders, with more trusted partners added later through what Google calls its Fairwind Program. Here's my problem with it. Per Dealroom, Google hasn't published a technical report or a parameter count, and Ars Technica notes Carbon could be a post-training upgrade of Argon rather than a new architecture. The alignment claims, like successful cheating on Terminal-Bench dropping to zero, are Google's own measurements. And nobody outside that defender group can build on it yet. There's no date for developers, and the two-dollar price is introductory before it doubles. I'd keep routing to Opus or Sol until there's a public API and a cost-per-task figure, and Google hasn't shared one.
Nathan Lambert, writing at Interconnects about the rebuilt Artificial Analysis index:
Open models’ demise from this seems greatly exaggerated, but I do think the previous index was flattering the open weight models too strongly.
Artificial Analysis launched version five of its Intelligence Index on Tuesday. Sixty percent of it now runs on private test sets or held-out answers. New in the mix is Cognition's FrontierCode 2.0, graded on whether a maintainer would merge the change. And that rewrites Monday's Kimi story. Under the old index, the best open model sat three points off the leader. On v5, Kimi K3.1 Pro scores forty-six, twelve behind Opus 5.5 at launch. Among models whose weights are actually out, DeepSeek V4.1 Flash and GLM-5.3 top out at thirty-two. Claude Haiku 5.5 beats both of them. Lambert's read is that v5 is built to measure where frontier models are heading, something like a general agentic collaborator, and open models are furthest behind on the science evals. He also flags the cost of going private: you trust the benchmark builder completely, and the community loses the chance to find the bugs. And he points out that Artificial Analysis makes its money from labs paying to verify their results. So the referee gets paid by the teams. The figure I'd actually use is cost. Among the top ten, cost per index task varies more than thirteen-fold, from ninety-six cents for GPT-6.1 Terra to thirteen dollars and four cents for Claude Fable 5.1. GPT-6.1 Sol lands within four points of Opus at about an eighth of the cost. Lambert expects the next open cycle, he names GLM 6 or Kimi 3.5 in two to three months, to claw ground back. Kimi K3.1 Pro's weights are still promised for November 4th.
TechCrunch, on Microsoft's results Wednesday:
Microsoft said it made $10 billion in revenue from commercial arrangements with OpenAI, including revenue-sharing payments, in the quarter ending September 30. It underscores the still-enormous financial dependency of the partnership.
Total revenue was ninety-two point three billion dollars, up nineteen percent, Microsoft's tenth straight record quarter by TechCrunch's count. Microsoft Cloud brought in sixty-three point nine billion. And for the first time, Microsoft put a dollar figure on Azure: thirty-three point two billion for the quarter. Ten billion dollars from one partner in one quarter. Last week OpenAI put its models on Google Cloud and made a point of saying Microsoft is still its primary cloud. Now we know what primary is worth. Alphabet reported Wednesday as well. Revenue was a hundred twenty-seven point six billion, up twenty-five percent, and Pichai says that puts the company on a half-trillion-dollar annual run rate, with AI investment pushing search query growth. The Bad Influence roundup, citing Rohan Paul, has Google Cloud revenue doubling to thirty point seven billion. That's within about two and a half billion of Azure. Emarketer's Nate Elliott told The Register the run-rate line was a sly nod to how the AI-only companies talk about revenue, and that it's almost ten times what either OpenAI or Anthropic claims. The incumbents have learned to talk like startups. Amazon reports today. On Monday we heard Jassy say AWS is largely sold out through 2027. Now we get to see what sold out looks like in the actual numbers.
Governor Kathy Hochul's office, announcing Wednesday's bill signing:
Violations would allow the attorney general to seek civil penalties of up to $25,000 per violation along with damages, restitution, the destruction of unlawfully collected data or any AI models trained on it, and any other appropriate injunctive relief.
The bill comes from State Senator Kristen Gonzalez and Assemblymember Alex Bores. It bans a list of chatbot features for anyone under eighteen, and for any user the company hasn't confirmed is an adult. No pretending to be a real person, or alive, or to have feelings. No claiming a romantic or authority-figure relationship with a child. No unprompted emotional questions, and no reusing a kid's health information from earlier conversations. Nothing that promotes self-harm, eating disorders or drug use, and nothing that encourages secrets or isolation. The builder part is age verification. Companies have to offer more than one way to verify age, at least one that doesn't depend only on a government ID, and they have to delete the age data afterward. IAPP says it takes effect January 1st, 2027. That's about two months to ship age checks, or strip those features for everyone. And look at the remedy in that quote. The attorney general can seek destruction of models trained on unlawfully collected data. That reaches past the app and into the weights. Justine Gluck at the Future of Privacy Forum told IAPP the prohibitions are broadly worded and somewhat undefined, and that it's unclear where ordinary conversational responsiveness ends and prohibited behavior begins. Every support bot that says it's sorry you're having trouble is about to get a legal review. IAPP counts twelve states with some chatbot safety law as of July. On October 12th we covered New York's RAISE Act for frontier models. This one goes after the consumer layer.
Microsoft, Alphabet and Amazon all report this week, and the spending under those revenue lines is where the data centers come from. Our sister show, The Data Center Daily, tracks that layer every day: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Search for The Data Center Daily and put it next to this one in your queue.
Quick recap: Carbon is first on the leaderboard and last in line for developers, the new index puts open weights twelve points back, Microsoft made ten billion off OpenAI in a single quarter, and New York wants proof of age before a chatbot gets friendly. Links to all of today's reporting are in the episode notes. AI Daily Briefing is one of the Lantern Podcasts. Talk to you tomorrow, Friday the 30th.