Google opens Gemini 4 Argon to developers and enterprises: a 1M-token output ceiling and limited-time API pricing of $2 per million input tokens, $10 per million output tokens

Start with the official timeline and the access channels. On September 30, 2026, Google published a post titled 'Gemini 4 Argon: our next era of frontier intelligence', unveiling the frontier model and stating that at that point it was rolling out only to trusted cyber defenders inside the Fairwind Program, while the company took part in the U.S. government's voluntary pre-release model access process. On October 13, 2026, Google published 'Gemini 4 Argon is now available for everyone', opening the model to developers, enterprises, scientists and consumers. The channels are specific: developers and enterprises get it through the API in AI Studio and the Gemini platform; Google AI Ultra subscribers get it in Antigravity, the agentic development environment; Ultra and Pro subscribers get it in the Gemini app and in AI Mode in Search. On pricing, the stated limited-time rate is $2 per million input tokens and $10 per million output tokens, with cached input tokens taking a further 95% discount off the input price. One number deserves its own line: Argon's output ceiling rises from 64K in the previous generation to 1 million tokens, which Google calls industry-leading. VentureBeat reported the same day that the model had previously been open only to a small group of early testers, with broad availability landing on October 13 — thirteen days after the September 30 unveiling.

The hardest part of the official post is what Google itself did with Argon, because those examples come with numbers. Google says more than 100,000 Google employees use Argon each month through an internal version of Antigravity, and that over a four-week period engineering teams increased agentic code submissions by 35% while cutting post-submission code rollbacks by 30%. The Chrome team used it to separate roughly 6 million lines of browser feature code inside Chromium's approximately 40-million-line codebase, compressing an architectural refactor expected to take months into weeks. The team behind Google's open-source Agent Development Kit for Python handed Argon a rules file plus a one-paragraph objective and let it work overnight; the result cut in-memory session preparation time by 40% for long-running sessions and sped up cold imports by 21%, with the changes already landed in the public adk-python repository. The open-source quantum chemistry package xTB got up to 2.3x faster at predicting molecular structures and 1.8x faster at conformer search. Google's open Gemma 4 E2B reached a 2.85x end-to-end inference speedup on an A100 via vLLM, going from roughly 37,900 to more than 108,000 tokens per second with 100% token parity and no quality drop. In quantum computing, Argon translated an internal version of the AlphaQubit error-correction decoder into a single highly optimised hardware kernel, nearly doubling inference throughput; for the arcsin subroutine used in quantum simulation, an implementation written by Argon and formally verified in Lean cut spacetime cost by more than 100x against a published quantum CORDIC baseline. In fusion, a Google team working with Professor Matt Landreman of the University of Maryland found a broader class of counterexamples to Grad's conjecture.

Science and cyber defence are the two areas the post showcases hardest. On rare disease research, scientists at Gamow Labs used Argon to analyse de-identified genomes from newborns with rare developmental lung disorders; in one case it pinpointed a chromosomal rearrangement that routine software had missed for years and explained why it went undetected. On private evaluations built from real clinical and lab data at Gamow Labs, Argon scored 57% on RareBench and 52.2% on LabBench. On enzyme design, Google scientists worked with Argon to design serine esterases, assembling data, tuning models and applying biochemistry-inspired filters within hours to reach five novel designs, with subsequent laboratory testing confirming catalytic activity. On robotics, Google says Argon scored 87% on the Embodied Reasoning benchmark, which tests understanding and problem solving in 3D physical environments, beating other frontier models; in an internal test Argon watched raw video of a human playing cornhole, mapped spatial coordinates, calculated whole-body joint movements and wrote the simulation code itself, then transferred it to Apptronik's Apollo humanoid, which sank seven consecutive shots in a real-world trial. On cyber defence, Google said it is expanding the Fairwind Program to hundreds of additional partners starting on launch day; Wiz used early access in two cases, identifying a gap that risked exposing sensitive government data and a public file download bypass that exposed a master key and could grant administrator access. Google is also giving 10,000 scientists free access to its highest Antigravity usage tier for six months, onboarded on a rolling basis.

A few boundaries are worth stating. First, Argon ranking first on the Vals Index is Google citing a leaderboard position; independent assessments are less uniform. VentureBeat obtained a review before the broad release from Sam Witteveen, CEO and co-founder of enterprise AI consultancy Red Dragon AI, whose words were that this is a really, really, really good model — but not, in his view, the best one out there. VentureBeat's own read is that Google's biggest achievement may be strategic: building a sufficiently capable general-purpose foundation for increasingly sophisticated products and enterprise workflows. Second, the word limited matters. Two dollars per million input tokens and ten dollars per million output tokens is a limited-time discounted price per Google's own post, not a long-term pricing commitment. Third, on safety, Google states it is engaged in the U.S. government's voluntary pre-release model access process and will keep iterating on guardrails as access expands; this announcement itself provides no customer-facing quotas, no enterprise migration schedule and no third-party head-to-head comparison with rivals. One more detail worth noting for readers: China is not on Argon's list of access channels, and the jump to a one-million-token output ceiling directly raises the ceiling for long-context tasks.