Google Releases Gemini 4 Argon: $2 per Million Tokens and First Place on the Vals Index

Google DeepMind has released Gemini 4 Argon, a flagship model for long, multi-step tasks. Since October 13, it has been available globally to developers and companies through the API in AI Studio and on the Gemini platform. Google AI Ultra subscribers get it in Antigravity, and Ultra and Pro subscribers also get it in the Gemini app and AI Mode in Search. Argon ranks first on the Vals Index, which measures model performance in finance, coding, legal work and taxes. Launch pricing is $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper.

A 6-million-line refactor and a Gemma speedup in a few days

The announcement puts its main emphasis on engineering work. Inside Google, Argon paired with Antigravity raised the number of agentic code submissions by more than 35% and cut rollbacks by almost a third.

The Chrome team separated the web page processing core from the legacy folder of browser features: about 6 million lines inside the 40-million-line Chromium codebase. The refactor spanned hundreds of files and would normally take months, but it was done in weeks. For users, this means live pages will be able to move between tabs, side panels and floating windows without reloading. A whole class of browser crashes will also disappear.

The Agent Development Kit team gave Argon a .agents/skills rules file and a one-paragraph goal: remove framework overhead and break nothing. The model worked autonomously overnight. The result is already in google/adk-python: in-memory session preparation in long conversations is 40% faster, and cold imports are 21% faster.

The most telling figure for infrastructure engineers involves the open Gemma 4 E2B. Running through vLLM on an A100, end-to-end inference sped up 2.85x: from ~37 900 to more than 108 000 tokens per second. The tokens matched 100%, and quality did not drop. The changes ranged from attention tuning to speculative decoding, and the whole job took a few days instead of months.

Science and security: from quantum kernels to enzymes

The showcase here is unusually broad:

• xTB: one of the most cited open-source quantum chemistry packages became up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search. The changes have been accepted into the official repository.
• AlphaQubit: Argon ported an internal version of the ML quantum error decoder into a single optimized hardware kernel and nearly doubled inference throughput.
• Quantum circuits: an arcsin implementation written by Argon and formally verified in Lean cuts spacetime cost by more than 100 times relative to the published quantum CORDIC baseline.
• Fusion: working with Professor Matt Landreman of the University of Maryland, the Google team found a broader class of counterexamples to Grad's 1967 conjecture on stellarators. This gives the community reference tests for simulators.
• Rare diseases: Gamow Labs used Argon to find a chromosomal rearrangement that routine software had missed for years. On private tests, the model scored 57% on RareBench and 52.2% on LabBench.
• Enzymes: in a matter of hours, Argon gathered data, tuned a model and filtered candidates down to five new serine esterases. Lab tests confirmed catalytic activity in all five.
• Isomorphic Labs: parsing complex PDFs and patents dropped from weeks to hours.

In cybersecurity, the Fairwind program is expanding, with hundreds of new partners joining. Wiz uses Argon in its "Scan for Good" initiative. The model helped close a gap that threatened sensitive data at a government agency. In medical infrastructure, it chained several vulnerabilities together and found a file-download bypass that leaked a master key and gave administrator access. Hundreds of thousands of patient records were at risk.

Tokens, pricing and who it suits

The 95% discount on cached input changes the economics of agents. Consider a loop (editorial calculation) of 50 calls with a 200 thousand-token context and 2 thousand tokens of output per call. Without caching, input alone costs about $20, plus $1 for output, for a total of ~$21. With caching, input costs $1, and the whole run comes to roughly $2. The longer the context and the more steps, the better the setup pays off with the repository in cache and reusable rules such as .agents/skills.

The practical takeaway for teams: Argon is worth trying wherever the result is verifiable. Tests, benchmarks, formal proofs and lab checks let you send an agent off overnight and get a ready pull request in the morning. The ADK, xTB and Gemma cases all follow that pattern. Google calls the price introductory, so it is wise to budget with a margin.

For researchers, Google is offering the top Antigravity tier free for 6 months to 10 000 people, with phased onboarding and selection. Applications are already open.