Good morning, it's Wednesday, October fourteenth.

In today's briefing we have Anthropic's research on small models replicating across edge devices, Anthropic swapping out git for its own source control, and Google opening Gemini 4 Argon worldwide with a price doubling on the way.

First up - Today in the big model news;

Google

Gemini 4 Argon, first released only to vetted cyber defenders, is now available through the AI Studio API, the Gemini app for Ultra and Pro subscribers, and Antigravity. There is no free tier. Pricing is two dollars in and ten dollars out per million tokens, and it doubles from December. Gizmodo puts GPT-6 Astra at ten and fifty dollars. Gizmodo also relays employee skepticism about Argon's coding ability, which Google is recorded as denying. A model that started behind access controls is now generally available, so frontier capability does not always stay gated. Costs are locked in only until the price step, so build model switching into your routing before then.

Anthropic

There is a new controversy about how dangerous small open-weight models are. Anthropic has published research on whether they could copy themselves across devices.

The finding is that models of thirty billion parameters or fewer can already perform parts of a self-replication chain. They often fail when they have to compose the steps end to end.

In one test, Z.ai's open-weight GLM-5.2 replicated across all eight nodes of a Google Cloud setup in four of five trials. Anthropic calls that an upper bound, because the exploit was deliberately trivial: the devices had unauthenticated Android debug access left open.

Anthropic projects small models reach that capability around the end of this year. It calls a full rogue replicator theoretical.

Now the contested part. Mozilla's Marco Figueroa accuses Anthropic of overstating the findings. The paper itself concedes most edge devices are not reachable from the web, and no mass compromise has been recorded. No independent assessment of the claims has surfaced.

The paper's data on the Mirai botnet shows where cheap defenses pay: ninety-two percent of infections cleared within fourteen days when subscribers were notified, against seventy-four percent without. Secure default credentials and egress monitoring are product requirements for anything shipping on consumer or IoT hardware.

In the harness, tools and orchestration world;

SpaceX Super Intelligence is bundling Cursor, Grok and X Premium into one subscription across five tiers, from free to two hundred dollars. The latest Grok models require the thirty dollar Premium tier. The company claims the comparable plan costs the same or less than X Premium Plus, and rollout to existing subscribers is gradual, so the pull is unproven. The bundle arrives as an upgrade prompt inside an app subscribers already open. ChatGPT and Claude own the assistant relationship but have no feed to bundle, so this tests whether distribution can beat a proven assistant habit. Cursor's standalone pricing is threatened only if bundled users change behavior.

In AI Infra

Anthropic has replaced git with an in-house source control system called Scion, plus a code review tool called Antique. Merges rose from about twenty-five hundred a day in May to over thirty thousand a day by September. Git assumes offline, distributed clients, while most of Anthropic's clients are always-online agents and jobs, and scaling git had meant several hundred internal mirrors. Anthropic reports a CPU and memory footprint roughly two hundred times smaller, checkouts dropping from minutes to seconds, and twenty-five percent more CI capacity, all self-reported. A seven-person team built it in five months, with Claude writing most of the code. Version control and CI were designed around human commit rates, so budget for them as the next bottleneck when scaling coding agents.

In other news…

Character.AI has filed for Chapter 11, with assets and liabilities of fifty to one hundred million dollars. Disney has agreed to buy some assets for twenty million dollars, plus a non-exclusive perpetual technology license. The platform keeps running during the sale, and CEO Colin Adams said it will be wound down if no bid allows continued operations. Cited headwinds include lawsuits, one over a fourteen-year-old's death, a Texas child-safety investigation, and the cost of monetizing heavy consumer use, for a company once valued above one billion dollars. Companion products carry inference costs and child-safety liability that can outrun funding, so both belong in the core risk model for anyone building in this category.

Quick hits from the consumer side;

Google Play now lets US adults build private apps in Google AI Studio and run them on their own Android device, hidden from search, with sharing planned in the months ahead.

That's the briefing. Have a great day, and don't forget to subscribe.