Google launches Gemini 4 Argon with aggressive pricing and record-breaking performance

Google released Gemini 4 Argon on Monday with immediate global availability for developers and enterprises. Access opens through the API in AI Studio and the Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The model targets deep work on complex, multi-step tasks in code, finance, legal, cyber and science, and ranks first on the Vals benchmark, which measures performance in critical economic sectors including finance, coding, legal work and taxation.

Introductory pricing undercuts the frontier tier: $2 per million input tokens and $10 per million output tokens — roughly 7.4 and 37 shekels respectively — with a 95% discount on cached input tokens. Google says the price is temporary but has not set an end date or a future target price.

Inside Google, Argon is already tackling the hardest problems. The Chrome team used it to separate the core that handles web pages from a legacy folder of roughly 6 million lines of code inside the Chromium base (40 million lines total). The result: an architectural refactor across hundreds of files finished in weeks instead of months, enabling live pages to move between tabs, side panels and floating windows without a reload and eliminating an entire class of browser crashes.

The Agent Development Kit (ADK) team gave Argon the framework rulebook and a single-paragraph goal: cut overhead without breaking existing applications. The model worked autonomously overnight and produced optimizations now merged into google/adk-python: a 40% reduction in in-memory session preparation time for long sessions, and a 21% speedup in cold import. In the xTB quantum chemistry package, one of the most cited in the field, Argon accelerated molecular structure prediction 2.3x and conformer search 1.8x. The changes were contributed back to the official GitHub repository.

On A100 GPUs via vLLM, Argon delivered a 2.85x end-to-end inference acceleration on Gemma 4 E2B, from 37,900 to more than 108,000 tokens per second, with full token parity and no quality loss, in a process that took days instead of months. Google also notes significant improvements in spatial and mathematical reasoning, though full details have not yet been published.