Analysis
Google released Gemini 4 Argon on Oct. 13, making its newest frontier model available immediately through the Gemini API to developers and enterprises, and through the Gemini app, AI Mode in Search and Google Antigravity to Google AI Ultra and Pro subscribers. The company is pricing API access at an introductory $2 per million input tokens and $10 per million output tokens -- a rate that holds only through Nov. 30, after which it doubles to $4 and $20. Google says Argon targets "deep work" on multi-step tasks in coding, finance, legal, cybersecurity and science, with a 1-million-token unified input-output context window.
How does Argon compare to rival frontier models?
Independent benchmarking from Artificial Analysis puts Argon's Intelligence Index score at 53, tying OpenAI's GPT-6 Astra and roughly a point ahead of GPT-6.1 Sol, but trailing Anthropic's Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56). That's a 23-point jump over Google's prior Gemini 3.1 Pro Preview, and Artificial Analysis also credits Argon with a 15% hallucination rate against 51% to 54% for the two GPT-6 variants it tested. The timing puts Argon directly against Meta's own frontier re-entry: Meta's Superintelligence Lab shipped Muse Wonder a day earlier at the same $4/$20 standard rate Argon will charge once its own discount expires, setting up a six-week window where enterprises can run both models at a discount before either hits list price.
“Whether Argon's edge holds past that window, and whether its 15% hallucination rate stands up outside Artificial Analysis's own test set, is the next thing to watch.”
Google's own benchmark claims go further: the company says Argon ranks No. 1 on the Vals Index, which covers finance, coding, legal and tax work, and scored 87% on the Embodied Reasoning benchmark, which it says beats every other frontier model tested. Those are Google-run evaluations, not third-party scores.
What Google says Argon can already do
The launch post leans on internal case studies rather than customer testimonials. Google says Argon helped refactor a roughly 6-million-line legacy folder in Chrome in weeks instead of months, cut cold-import time 21% and in-memory session prep time 40% for its Agent Development Kit, and sped up molecular structure prediction up to 2.3x for the quantum-chemistry tool xTB. In a robotics demo, a policy trained on an Argon-generated throwing motion sank seven straight cornhole shots on Apptronik's Apollo humanoid. On security, Google says Wiz used early access to find and patch a vulnerability exposing healthcare software, and it's expanding its Fairwind cyber-defender program to "hundreds" more partners with a version of Argon that has its cyber guardrails removed for vetted security researchers.
On safety, Google says adversarial harmful-request violations fell nearly 2x versus prior Gemini models and that prompt-injection attacks succeeded only 0.7% of the time across 15 adaptive attempts on the Gray Swan benchmark -- numbers aimed squarely at the enterprise security teams that have slow-walked agentic AI rollouts over exactly those failure modes.
What the headline misses: every one of those improvement figures is Google's own measurement, run on Google's own test suite, with no outside party checking the methodology. The one genuinely independent number available -- Artificial Analysis's Intelligence Index -- shows Argon tied with OpenAI's best model and still behind Anthropic's, not leading the field outright. And the pricing that makes Argon look cheap today is temporary by design: any enterprise locking in a vendor contract this month should get that vendor's own eval numbers against Claude and GPT-6, because the rate resets higher on Dec. 1 regardless of whose benchmark wins.
Google is betting that six weeks of discounted access converts into locked-in enterprise and developer usage before OpenAI or Anthropic can respond with a release of their own. Whether Argon's edge holds past that window, and whether its 15% hallucination rate stands up outside Artificial Analysis's own test set, is the next thing to watch.