Gemini 4 Argon is now generally available, and for anyone running an agent on a Windows machine the headline number is the price: $2 per million input tokens and $10 per million output tokens as an introductory rate, with cached input tokens at 95% off. Google published the model in its own announcement and The Verge covered the launch on October 13, 2026 — Google’s first major frontier release since Gemini 3 in late 2025.
Key facts
| Fact | Detail |
|---|---|
| Question answered | Gemini 4 Argon Is Out: Google’s First Frontier Model Since Gemini 3 |
| Short answer | Google released Gemini 4 Argon on October 13, 2026, its first major frontier model since Gemini 3 in late 2025. |
| Reading time | ~3 min |
| Last updated | 2026-10-14 |
Where Argon is actually switched on
Google’s post is specific about availability, and it is not universal:
- Developers and enterprises get it through the API in AI Studio and the Gemini platform.
- Google AI Ultra subscribers get it inside Google Antigravity, the company’s agentic development platform.
- Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search.
The Verge noted that it is not clear when, or whether, Plus subscribers and free users get access. For desktop agent work, the API is the practical route today.
The claims, and who is making them
Everything about Argon’s superiority currently comes from Google or from reporters restating Google. The company says Argon excels at deep work in coding, finance, legal, cybersecurity and science, and that it ranks #1 on the Vals Index, which measures performance across sectors such as finance, coding, legal and tax work. The Verge adds that Google claims Argon outperforms GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 on benchmarks spanning coding, cybersecurity, science, finance and law, and that it hedges, repeats itself and reaches for cliches less often.
None of that has been independently rerun, so treat the ranking as a vendor claim with a press echo rather than a measurement you can bank on.
The internal results Google published
The most interesting part of the announcement is the list of jobs Google says Argon already did inside the company:
- In Antigravity, Google says teams increased agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.
- On Chrome, Argon helped separate core webpage handling from a roughly 6-million-line legacy folder inside Chromium’s 40-million-line codebase, finishing an architectural refactor across hundreds of files — months of work compressed into weeks. The user-facing payoff: pages moving between tabs, side panels and floating windows without reloading.
- On the open-source Agent Development Kit for Python, Argon was handed a rulebook and a one-paragraph goal, then worked overnight: in-memory session prep for long sessions fell 40% and cold imports got 21% faster, changes now live in google/adk-python.
- On xTB, a widely cited open-source quantum chemistry package, Google reports up to 2.3x faster prediction of molecular structures and 1.8x faster conformer search, contributed to the upstream repository.
- On Gemma 4 E2B, a 2.85x end-to-end inference speedup on an A100 through vLLM, taking throughput from about 37,900 to over 108,000 tokens per second with 100% token parity.
The Verge adds that Google is also using Argon to find and patch cybersecurity vulnerabilities, to halve the compute behind YouTube’s comment section, and to debug the physics simulator for Alphabet’s drone delivery unit Wing.
What this changes for a desktop agent setup
Three things are worth acting on this week. Cheap input tokens matter more than cheap output tokens for agents that re-read large files, and cached input at 95% off is the line item that decides whether long, repetitive runs stay affordable. The introductory price doubles at the end of November, so any cost model you build today should be re-checked in December. And model-agnostic agent runners can now add Argon without a rewrite, which is the argument for keeping your agent separate from any single vendor’s IDE.
A short checklist
- Pick one recurring task you already run an agent on and repeat it with Argon through the API while the introductory price holds.
- Log token counts per run, not just wall-clock time — the pricing advantage lives on the input side.
- Keep a fallback model configured, because Plus and free-tier access is still unannounced.
If you would rather have an agent that runs on your own Windows desktop, with your files staying where they are, download MOOGH and point it at the model you trust most.
Related reading: