Google’s best model is now just something you can call. Gemini 4 Argon was announced on September 30 for a small circle of cyber defenders; on Tuesday it reached the public Gemini API at $2 per million input tokens and $10 per million output, half of Claude Opus 5.5’s price and a fifth of Fable 5.1’s. The catch is in Simon Willison’s fine print: that price doubles on December 1, which would put it level with Opus. Google also pushed a robotics write-up claiming 87% on the ERQA reasoning benchmark. These are Google’s own numbers, and the vendor-chosen benchmarks lean its way, but a cheap model that runs the benchmark suite for $2,407 is a pricing story as much as a capability one. Enjoy the intro rate.
The bottleneck moved
The most telling post of the day is not a model launch. Anthropic explains why it replaced git. Merges to its monorepo went from about 2,500 a day in May to over 30,000 by September, mostly because Claude writes code faster than humans review it. Git assumes offline clients holding a full local copy. Anthropic’s clients are almost all always-on agents and CI jobs, and about 25 million checkouts a day had forced it to run hundreds of git mirrors. So seven engineers built Scion, a server-centric system that, per the company, uses roughly 200 times less CPU and memory than the mirror setup. Willison noted it looks a lot like Subversion, which is either a delicious irony or a reminder that old designs were fine for always-online machines.
Read that next to the usual debate about whether AI makes developers faster. The interesting part is what breaks next: version control, code review, CI. Infrastructure built for human cadence is starting to creak, and the fixes look like unfashionable centralization.
Humans, doing human things
Not all of today’s trouble involves models. An OpenAI engineer, Abdulrahman Alfozan, was arrested after a recording device was found in an executive conference room, hidden in toilet paper and plugged into a TV’s USB port. Prosecutors allege he used an internal productivity tool to look up meeting rooms, and that a chatbot account tied to his spouse asked how to hide a device. OpenAI says it found no sign that weights, source or customer data were touched. That is an allegation, with a detention hearing set for October 19.
And in France, attackers pushed supermarket refrigeration to 50°C overnight, hitting around 150 Intermarché stores plus Cooperative U and Carrefour outlets, with fires and discarded stock. Danfoss says the affected units ran outdated firmware, and French authorities are investigating, with no suspects named. No frontier model required: unpatched industrial controllers on the internet remain the cheapest weapon around.
