Google Releases Gemini 4 Argon, Its First Major Frontier Model Since Gemini 3
Gemini 4 Argon is rolling out globally to the paid Gemini API, Pro and Ultra subscribers, and Antigravity, at an introductory $2/$10 per million tokens. It ranks #1 on Vals AI’s independent Vals Index, though Google’s own table shows it trailing on some coding and science tests.
By 1EyeAI Desk ·
Correction (Oct. 13, 2026, 8:05 PM ET): An earlier version of this story labeled Argon’s #1 ranking on the Vals Index as a Google claim that hadn’t been independently verified. Vals AI, which runs that index, lists Argon first on its own leaderboard (68.90%), so we’ve relabeled it and added Vals’ figures, which show Claude Sonnet 5.5 (67.04%) in second place. See our corrections log.
Google announced Gemini 4 Argon on Tuesday (Oct. 13, 2026, about 12 PM ET). Developers can use it through the paid Gemini API in AI Studio and Android Studio, enterprises through the Gemini platform, Google AI Ultra subscribers in the Antigravity coding platform, and Pro and Ultra subscribers in the Gemini app and AI Mode in Search. The Verge notes it’s unclear when, or whether, free users will get it.
Pricing (official): an introductory $2 per million input tokens and $10 per million output tokens, with cached input 95% off, through Nov. 30, 2026. After that it rises to $4 and $20. Through the API, Argon has a 1M-token limit shared between input and output.
Independently checked: Vals AI’s own Vals Index leaderboard (updated Oct. 13) ranks Argon #1 at 68.90%, ahead of Claude Sonnet 5.5 (67.04%) and Claude Opus 5.5 (66.97%), with GPT-6 Astra at 63.13%. Company claims (not independently verified): Google says Argon also leads on several legal, finance, long-context and multimodal tests. Its own benchmark table also shows Argon trailing on some coding and science tests: Terminal-bench 4.0 (57.4% vs. Opus 5.5’s 66.4%), FrontierSWE v2 and Terminal-Bench Science (both led by GPT-6 Astra). Google also says Argon scores 87% on the ERQA embodied-reasoning test and has made agentic code submissions inside Google more than 35% higher.
On science, Google says Argon helped researchers find a broader class of counterexamples to “Grad’s conjecture,” a 1967 hypothesis about stellarator fusion magnetic fields, working with Prof. Matt Landreman (University of Maryland). It also says Argon nearly doubled the throughput of its AlphaQubit quantum error-correction decoder and designed enzymes that showed catalytic activity in lab tests. For robotics, Google’s developer post describes Argon turning a video of a person playing cornhole into simulation code that an Apptronik Apollo humanoid then used to sink seven straight shots.
Safety: Google says external bio red-teamers unlocked fewer than 1% of the bio sub-tasks it tests for, and that vetted cyber defenders in its Fairwind Program can apply for access without cyber guardrails. Details are in the model card.
What this means for you: If you pay for Gemini Pro or Ultra, Argon should show up in the app today. If you build on APIs, run your real workloads through it before the Nov. 30 price jump. On paper, Argon costs about a fifth as much as GPT-6 Astra (as reported by The Verge), but cost per token isn’t cost per task. Vals’ independent leaderboard backs up the top overall ranking, but treat Google’s other benchmark wins as its own claims until independent tests confirm them.