The frontier model Google kept for defenders goes on sale
Edition of 2026-10-14
# AI For All ### Wednesday, October 14, 2026
As of today, a financial adviser, a lawyer or a tax manager with a Google subscription can put the company's strongest model on client work, at a published price instead of by invitation.
The model Google kept for a chosen few is now on sale to anyone
Google put Gemini 4 Argon, its strongest model, on general sale worldwide. Until this week it went only to a closed circle of security professionals. Developers reach it through the API and in AI Studio. Subscribers reach it in the Gemini app and in the AI Mode of Search. Google published an introductory price of *$2 per million input tokens*, with cached input discounted.
The ranking is the part to read twice. Google says Argon now stands first on the Vals Index, which scores models on finance, coding, legal and tax work. The capability figure it leads with is physical reasoning, where the model reports *87%*, above every other frontier model. Those are the vendor's own results, and nothing published so far prices the model against a rival on the same job.
Google is also buying goodwill in research: it is offering *10,000 scientists* free use of its highest tier in Antigravity for six months.
The chatbot company that defined a category is selling itself in court
Character Technologies, which runs Character.ai, filed for Chapter 11 protection in Delaware to sell all or most of its assets. Before the filing it agreed that Disney Worldwide Services, an affiliate of the Walt Disney Company, would open the bidding for a non-exclusive perpetual licence of some of its intellectual property.
Disney's offer is *$20 million* and a loan of staff: the studio has hired former employees of the company so that the platform keeps running through the sale. That is a turn. Disney had publicly attacked the company over its characters a year earlier. Now it is buying rights to the machinery.
For anyone paying for the service, the dates are what matter. The company expects a sale to close in late November or early December. Subscriptions keep billing as normal. If no bid keeps the platform alive, the stated plan is to wind it down.
A model small enough for a laptop now holds its own at using tools
Liquid AI published an open-weight model built to run agents on the machine in front of you rather than in a data centre. Both the model and its base version are downloadable from Hugging Face today, with support in the usual local tools.
It is a reasoning model, so it works through a problem before answering, and it does that at a fraction of its own size per step. That is what lets it fit on ordinary hardware. Liquid says it decodes *1.7x faster* than the next-best model in its class, using the least memory of the group.
The evaluation is broad rather than selective: *16 benchmarks* covering tool use, agentic workflows, knowledge and coding. Liquid says the model leads or matches much larger rivals on five of the eight tool-use tests, and trails them on knowledge-heavy and coding work, where it recommends pairing the model with search.
Weak signal
Google DeepMind and Singapore's Government Technology Agency have put a scam detector on the phone itself. The model is tuned on local scam conversations and runs on the device, so messages do not have to be sent away to be checked, which matters because the conversations people are tricked in are usually end-to-end encrypted. Singapore makes a good test bed: DeepMind says scams are more than half of the crimes reported there, and the losses run into the billions, with social engineering behind most of them. It stays a signal rather than a story because nothing in it is available to a reader today, and no messaging service has said it has integrated it.
The discount on the frontier will not last. Ask what the list price buys once the introductory rate is gone.