Gemini 4 Argon is out, and its price doubles on December 1

Google began rolling out Gemini 4 Argon on October 13. It is now in the Gemini app, AI Mode in Search, and the Antigravity coding tool. Before that it ran for vetted cyber defenders under the Fairwind program announced September 30. Argon is Google’s first frontier release since Gemini 3. Ars Technica The Verge

Pricing comes in two stages. Right now it is $2 per million input tokens and $10 per million output tokens, with 95 percent off cached input. That is an introductory rate, and it holds until November 30. On December 1 it doubles to $4 and $20. The output limit also grew from 64K tokens to 1 million, so a single call can now emit something the size of a repository rather than a few files. Google

The numbers Google leads with are 77.9 percent on DeepSWE v1.1, 91.7 percent on LVBench, and first place on Zapier’s AutomationBench at 51.3 percent. All three are Google’s own measurements.

The $2 rate ends December 1, and the coding tools need the $99 tier

Two things matter if you are building a product on a frontier model. The first is that those token prices are a launch rate, not a list rate. Model your unit economics at $2 and $10, and December doubles the input side of the bill. Argon stays cheap for a frontier model even after that: The Verge puts the launch price at roughly a fifth of GPT-6 Astra and Claude Fable 5, and still below its rivals per token after the increase. It also notes that a low token price does not guarantee a lower cost per task. Models burn different numbers of tokens to finish the same job.

The second is that where you use Argon changes what you pay for. In the Gemini app and Search, it goes to AI Pro at $20 a month and AI Ultra from $99. In Antigravity and AI Studio it is Ultra only. If your coding workflow assumes the subscription covers the model everywhere, that assumption breaks at the agentic surfaces. Calling Argon straight through the API is a separate path.

Some things are still open. Google says it will keep Gemini 3.1 Pro available for anyone who needs more than the Flash models, and it has not said when Argon reaches the lower Antigravity tiers.

The rest of today’s news

  • JPMorgan blocks outside AI agents from client accounts: Jamie Dimon said on the bank’s third-quarter earnings call that outside agents such as Meta’s Muse will not act inside customer accounts. Balances can still be read through open banking; moving money cannot. CNBC
  • Anthropic replaces Git with in-house Scion: daily monorepo merges rose from around 2,500 in May to over 30,000 in September, so seven engineers spent five months building Scion and migrated the whole monorepo, claiming one two-hundredth the CPU and memory footprint of its Git mirror setup. Anthropic
  • Reuters: OpenAI and Anthropic sought pre-publication control over evaluator reports: through October 5, both labs pushed for substantial say over what outside evaluators could disclose, with proposals including on-site lab monitors and bans on working with competitors. The Verge
  • Character.AI files for bankruptcy, Disney bids $20 million: the company filed Chapter 11 in Delaware on October 12, listing assets and liabilities of $50 million to $100 million, with Disney as the stalking horse for some intellectual property. NBC News
  • Epic Games shifts a Disney game to AI-first development: developers were directed to use tools like Claude for code and AI workflows for art, leaving people in a supervising role. The Verge
  • Decagon holds early buyout talks with ServiceNow: the enterprise AI support agent, valued at $4.5 billion earlier this year, has hired an investment bank and is in early sale discussions. Business Insider
  • Pluto Security exits stealth with $42M: the startup monitors enterprise AI agents and employee-built AI apps, and has raised an $8.5 million seed plus a $34 million Series A. TechCrunch
  • Meta extends ‘Your algorithm’ to Facebook: the natural-language topic controls that began on Instagram now cover the Facebook Feed and Reels. Meta