Google released Gemini 4 Argon on Oct. 13 to developers, businesses and paying consumers. Koray Kavukcuoglu, Google’s chief AI architect, wrote in the launch post that it ranks first on the Vals Index, which measures model performance on work in finance, coding, legal and tax.
Developers can use Argon through the API in AI Studio and the Gemini platform. Google AI Ultra subscribers get it in Google Antigravity, Google’s coding tool, and Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search.
Argon launched at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens 95% off. Google’s post didn’t say how long the introductory price lasts.
What’s under the hood
Google’s model card describes Argon as a large sparse mixture-of-experts model, which only activates part of its network for each piece of input. It says Argon is a new model, “not a modification or a fine-tune of a prior model.” It takes text, images, audio and video, with a combined input and output window of up to 1 million tokens, and was trained on Google’s own TPU chips.
Google says it has used Argon on its hardest engineering problems. In Antigravity, it says, Argon helped teams increase agentic code submissions by more than 35% while cutting code rollbacks by nearly a third.
In one example, Google says Argon helped the Chrome team separate core webpage handling from a legacy folder of about 6 million lines of code inside Chromium’s 40-million-line codebase, finishing months of work in weeks.
In another, the model worked overnight on Google’s open-source Agent Development Kit and found changes that cut session preparation time by 40% in long sessions, according to the post. Google also says Argon sped up inference on its open Gemma 4 E2B model by 2.85 times. These are Google’s own accounts.
How outside testers rate it
Vals AI, which runs the index Google cited, evaluated Argon before launch. In a Sept. 30 update, Vals said Argon placed first of 41 models on the Vals Index with 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%. It also cost less per test, at $15.68 against $21.34 and $32.14.
The rankings have since shifted as newer models arrived. As of Oct. 31, Vals’ model page lists Argon fourth of 32 models on the index, at 69.24%. It ranks first on HLAB-Vals and the IOI programming test, and eighth of 59 on Terminal-Bench 4.0, a test of command-line work.
The ARC Prize Foundation, which runs the ARC-AGI reasoning tests, reported verified scores of 98.5% on ARC-AGI-1 and 91.3% on ARC-AGI-2 at high reasoning. On the newer ARC-AGI-3, Argon reached 49.8% with the standard test harness. With a harness that keeps the model’s reasoning state between requests, it reached 97.5%, the foundation said.
ARC Prize noted that API problems kept it from completing every ARC-AGI-3 run with the standard harness. It used Google’s introductory pricing for its cost figures.
Google says Argon writes more directly than earlier Gemini models, with “less hedging,” and follows style instructions better. Kavukcuoglu’s post also describes use at Isomorphic Labs, the Alphabet drug discovery company, where teams use Argon to work through complex PDFs and patents.
Source: Google Blog, Koray Kavukcuoglu, “Gemini 4 Argon is now available for everyone”
Additional sources: Vals AI, “Gemini 4 Argon Benchmarks, Cost and Capabilities”; ARC Prize, ARC Prize Foundation, “Gemini 4 Argon – ARC-AGI Results”; Google DeepMind, “Gemini 4 Argon – Model Card”
