Two dollars in, ten dollars out. That's what Google is charging per million tokens for Gemini 4 Argon until November 30, which is exactly what Claude Sonnet 5.5 and GPT-6.1 Sol cost at list price. On December 1 the price doubles to $4 / $20, which puts it level with Opus 5.5. The message is clear: try us now, while it's cheap, and see if you still want to leave in December.
Everything else I read today was also about price, just from the other end. Character.AI, which Google once paid $2.7 billion to license, is selling assets to Disney for $20 million. And dead startups are getting a few thousand dollars for their Slack archives.
Google finally ships Argon, and the price is the pitch#
Argon was announced on September 30 for a small group of approved security defenders. As of yesterday it's broadly available: in the paid Gemini API, AI Studio, Antigravity, Gemini Enterprise, and in the Gemini app for AI Pro and Ultra subscribers. VentureBeat has the full breakdown, and Ars Technica points out the catch: it's paid-only, and lower tiers don't get full access. If you're an individual and want it in Antigravity, you need Ultra.
The independent numbers are good, but they aren't a clean sweep:
Vals Index: 68.90%, ranked first, just ahead of Sonnet 5.5 (67.04%) and Opus 5.5 (66.97%).
Artificial Analysis Intelligence Index: 53, behind Opus 5.5 (58) and Sonnet 5.5 (56), and tied with GPT-6 Astra.
Terminal-Bench 4.0, from Google's own table: Opus 5.5 leads at 66.4% vs. Argon's 57.4%.
Output ceiling: up to 1 million tokens through a continuation protocol, up from 64,000.
Most of the rest is Google grading its own homework. It claims a 19.6% on Harvey's legal agent benchmark against Opus 5.5's 3.8%, and internal results like 35% more agentic code submissions and 30% fewer rollbacks across 100,000+ employees. Those are interesting. They are not evidence.
The part I'd actually test is the 1M-token output limit. A model that can rewrite a large module in one continued response changes how you design refactoring agents, if it stays coherent. Google says Argon helped reorganise about 6 million lines of Chromium feature code. I'd like to see that hold up on a codebase Google didn't write.
But be careful with the promo window. A VentureBeat survey out the same day found that 62% of companies swapped a model in a production app last year, and 59% of those saw performance drop and had to rework or revert. Only 19% said the swap was close to free. Six weeks at half price is a great time to run evals. It's a bad reason to migrate.
What a dead AI company is worth#
Character.AI filed for Chapter 11 about five years after it was founded, with assets and liabilities between $50 million and $100 million. Disney will pay $20 million for some assets plus a non-exclusive, perpetual license to its technology, and former CEO Karandeep Singh Anand is already at Disney. As Ars notes, that's the same Disney that sent Character.AI a cease-and-desist in 2025 accusing it of weaponising its characters.
The arc is brutal. Twenty million monthly users in October 2024. A $2.7 billion licensing deal that took the founders back to Google. Then lawsuits after a teenager's suicide, a Texas investigation, a ban on minors, and a pivot to microdramas to keep engagement up. CEO Colin Adams says that if no viable bid comes in, "the platform will be wound down."
Here's how I read it: the value walked out the door with the 2024 license. What's left is a toolkit for character chat, and Disney is the one buyer that owns characters worth chatting with.
MIT Technology Review covers the smaller version of the same market. AI labs are buying data from shut-down startups: Slack, Notion, Gmail and Figma workspaces, which become training material and RL environments for agents. One founder sold his crypto startup's codebase for about $25,000, with the workspace data expected to fetch about $15,000 more. Google paid $10 million for Spirit Airlines' data, including 100 million emails and 500 million Teams messages. The broker SimpleClosure says 47% of its deals now include a code or data sale.
"Legally radioactive." — attorney Greg Ewing, describing this kind of data, to MIT Technology Review
He's right. Calendars hold medical appointments and school pickups, the proceeds go to creditors rather than the employees who wrote those messages, and no US federal law stops the sale. If you're running a startup, assume your internal chat is a future asset on someone's balance sheet, and write like it.
Agents are getting a wallet with a leash#
Two commerce moves landed together. Shopify now lets shoppers connect any AI agent to buy from its catalog of over a billion listings, with no partnership integration required. Until now that meant Meta's Muse. And American Express introduced an Intent Intelligence Framework: a Create Intent API captures the merchant, budget and product preferences, the customer approves it, and the agent receives a scoped payment credential instead of the card number. Ask for white Nike sneakers in size 9 and an agent that tries to buy Adidas gets blocked.
This is the right design, because it puts the constraint at the payment rail and not in the prompt. Still missing: who pays when it goes wrong. Amex hasn't said how liability is split.
In brief#
Google Play now hosts private vibe-coded apps: build something in AI Studio and install it on your own Android phone through Play, without a public listing. It's US-only and 18+ for now. Sharing with friends and family is coming, and an app that gets popular can move to the public store after review. It's a small change, but it removes the reason so many hobby projects ended up as web apps.
My bet for December 1: Argon's price doubles, most teams that ran honest evals stay put, and the ones who switched on the strength of a launch-week benchmark table become part of next year's 59%.