On 13 October Google DeepMind made Gemini 4 Argon globally available. On the independent Vals Index it leads with 68.90%, at an introductory $2 per million input tokens.
Key takeaways
- Introductory price: $2 per million input tokens, $10 per million output tokens, cached input 95% off
- Vals Index, 13 October: 68.90% for Argon, 67.04% for Claude Sonnet 5.5, 63.13% for GPT-6 Astra
- In Google Antigravity, agentic code submissions rose over 35% and rollbacks fell by nearly a third
- The Chrome refactor covered hundreds of files inside the Chromium codebase
- Gemma 4 E2B: 2.85× faster inference on an A100 via vLLM, from 37,900 to over 108,000 tokens per second
A benchmark that measures work, not puzzles
The model is available globally through the API in Google AI Studio, in Google Antigravity for AI Ultra subscribers, and in the Gemini app and AI Mode in Search.
The Vals AI index measures performance across economically significant sectors — finance, coding, legal work and tax. That makes the Vals Index a measure of work rather than of academic puzzles.
| Model | Vals Index score |
|---|---|
| Gemini 4 Argon | 68.90% |
| Claude Sonnet 5.5 | 67.04% |
| Wonder Full | 64.57% |
| GPT-6 Astra | 63.13% |
| MiMo V2.6 Pro | 55.20% |
The margin over the nearest rival, Claude Sonnet 5.5, is under two points, while GPT-6 Astra trails by almost six. Which model actually wins therefore depends on the task, not the table position.
Introductory pricing, per million tokens:
What Argon did inside Google's code
With the model's help the Chrome team separated core webpage handling from 6 million lines of legacy browser features inside Chromium's 40-million-line codebase. Google DeepMind says the refactor covered hundreds of files and compressed months of work into weeks, letting live webpages move between tabs and panels without reloading.
Google DeepMind's clip on maintaining the vast Chromium codebase with Gemini 4 — the same refactor the company's blog describes.
In the Agent Development Kit the model worked autonomously overnight and cut in-memory session prep time by 40% and cold imports by 21%. On Gemma 4 E2B it reached a 2.85× inference speedup on an A100 via vLLM, with full token parity.
Science and cybersecurity
Argon translated an internal version of the AlphaQubit quantum error correction decoder into a single optimised hardware kernel, nearly doubling throughput. Its arcsin implementation, formally verified in Lean, cuts spacetime cost?spacetime cost: A resource measure in quantum computing: the number of qubits multiplied by how long the circuit runs. Lower means the subroutine is cheaper to execute. by more than 100× against a published CORDIC baseline.
At Gamow Labs the model pinpointed a chromosomal rearrangement routine software had missed for years, scoring 57% on RareBench and 52.2% on LabBench. Wiz uses it in its free Scan for Good initiative.
Why it matters
Most new-model announcements rest on academic benchmarks. Here Google shows a Chromium refactor, a speedup of its own framework and an optimised quantum decoder. The problem is verification: all those numbers come from the model's owner, and only the Vals Index result is independently confirmed.
What's next
- Expansion to hundreds more Fairwind Program partners starts now
- Isomorphic Labs uses the model to synthesise PDFs and patents in drug discovery work
- The $2 and $10 per million token prices are labelled introductory
