Google released Gemini 4 Argon on October 13. Sources for this post: an American Bazaar article citing Reuters, Online Tech Tips and a hands-on write-up from Computing for Geeks. I haven’t run the model. Where a number is Google’s, I say so.

What it is and where it runs

Google markets Argon as its flagship for long, multistep work: software engineering, finance, law, cybersecurity and science. Availability, per the coverage:

  • Gemini API: paid tier only. Computing for Geeks reports no free tier (its free key returned a quota of 0), plus availability on Vertex AI and OpenRouter on launch day.
  • Gemini app and AI Mode: Google AI Pro and Ultra subscribers.
  • Antigravity (Google’s coding tool): Ultra and eligible enterprise plans, with rollout beginning October 13 in the desktop app and CLI.

Pricing

Introductory API pricing is $2 per million input tokens and $10 per million output tokens. Per Reuters and the hands-on write-up, the rates double to $4 and $20 after November 30. Cached input gets a 95% discount on the input price, per Online Tech Tips. Budget for the post-promotion price, not the launch one.

Benchmarks: Google’s table vs independent scores

Google’s own table (compared against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5):

  • Argon leads several rows, including DeepSWE v1.1 at 77.9% and the Vals Index at 68.9%.
  • It trails Opus 5.5 on Terminal-Bench 4.0 (57.4% vs 66.4%), and other models lead on FrontierSWE v2 and OSWorld 2.0.
  • Google says it used single attempts at the highest thinking settings, and competitor scores mostly come from provider reports or public leaderboards, not one lab testing everything under identical conditions. One write-up also notes the table omits GPT-6.1 Sol and Claude Sonnet 5.5.

Independent (Artificial Analysis): an Intelligence Index of 53, tied with GPT-6 Astra, and a reported 15% hallucination rate on its AA-Omniscience test. On Terminal Bench 4, it scores 57%, behind Sonnet 5.5 and Opus 5.5.

A small hands-on test

The Computing for Geeks author ran three DevOps tasks three times each. Bash scripts and OpenTofu configs were fine for every model. But Argon’s twelve Kubernetes manifests passed linting and then crash-looped on k3s (a permissions error under a non-root user), working only after the author fed it the error. Opus 5.5 and GPT-6.1 Sol deployed correctly on all three runs. It’s a tiny sample, so I’d call it a warning, not a verdict.

Gotchas and things that don’t match

  • Context numbers differ between sources. One lists Google highlighting a one-million-token output limit. The hands-on write-up reports a 1,048,576-token context window with a 65,536-token output limit per synchronous request. Check the model card before designing around either.
  • Migration: on Vertex, thinkingBudget and the MINIMAL level return HTTP 400, explicit context caching is unsupported, and temperature, top_p and top_k are silently ignored.
  • Google’s productivity claim that internal teams using Argon through Antigravity increased AI-assisted code submissions by more than 35% and cut rollbacks by nearly a third is Google’s own and unverified.

Reuters, via American Bazaar, also reports that Google dropped its Gemini 3.5 Pro plans after delays and that Demis Hassabis stepped aside as DeepMind CEO. I couldn’t verify that independently, so read it as reported.

Should you switch?

For long-context work at the intro price, it’s worth a trial. For agent tasks that touch real infrastructure, keep tests in the loop. I put Argon next to this week’s other launches in the model comparison post, and I covered Google’s agent side in the Gemini universal agent post.

Sources: American Bazaar, Online Tech Tips, Computing for Geeks.