The Pulse
Google Opens Gemini 4 Argon to Developers and Businesses
Google made Gemini 4 Argon available to enterprise customers and developers on October 13 through Google Cloud, Google AI Studio and Antigravity. The model launches with introductory API pricing of $2 per million input tokens and $10 per mi
AI.info Team · Written with AI ·
Google opened Gemini 4 Argon to enterprise customers and developers on October 13, moving the model beyond its initial circle of cybersecurity testers. The company says customers can use it in Gemini for business, through APIs on Google Cloud and Google AI Studio, and in Google Antigravity. The release turns Argon from a limited-access model into a paid option for building software and business applications.
Argon reaches Google Cloud and developer platforms
Google lists the API model ID as gemini-4-argon. Enterprise developers can access it through the Gemini platform, Google AI Studio and Antigravity; businesses can also use it in Gemini for business on web, desktop and mobile. Google’s cloud announcement does not describe the rollout as a free public API release.
Introductory API rates run through November 30, 2026: $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount, Google says. After November 30, the listed standard rates rise to $4 per million input tokens and $20 per million output tokens.
A long context window, with a conditional million-token output
Google describes Argon as a reasoning model for complex software engineering, scientific and mathematical problem-solving, and multi-step agents. Its developer documentation lists a combined input-and-output context window of 1,048,576 tokens. The default maximum output is 65,536 tokens; reaching up to 1,048,576 output tokens requires the long-decode option.
The model accepts text, images, video, audio and PDF inputs, and returns text. Its documented tools include code execution, function calling, file search and search grounding; computer use is marked as a preview. Image generation, audio generation and the Live API are not supported, according to the same documentation.
Google’s evaluations show strengths and gaps
Google positions Argon for extended tasks that may involve reviewing large documents or making changes across a codebase. Its model card reports a 77.9% score on DeepSWE v1.1, an agentic software-engineering benchmark, and 68% on CWE-bench v1 for vulnerability remediation. Google says that CWE-bench result ties for first place among the models it compared.
The results are not uniformly highest across the company’s listed benchmarks. In the model card, Argon scores 55.0% on FrontierSWE v2, below the listed GPT-6 Astra score of 65.5%, and 57.4% on Terminal-bench 4.0, below Claude Opus 5.5’s 66.4%. Google also lists hallucinations, occasional slowness and timeouts among the model’s limitations, and says its knowledge cutoff is March 2026.
From Fairwind testers to paid access
Google first announced Argon on October 1, initially making it available to selected cybersecurity defenders through the Fairwind program while it gathered feedback and refined safeguards. The October 13 release expands access to developers and businesses, while Google says a limited group of vetted government, critical-infrastructure and cybersecurity partners can still receive access without cyber guardrails through Fairwind.
Early-access customers describe uses spanning software development and legal work. “As software development becomes more agentic, teams need models that can support long-running, multi-step work across all stages of the development lifecycle,” Stuart Moncada, GitLab’s vice president, says in Google’s announcement. The wider release now makes Argon available through named developer services at stated per-token rates, while leaving its documented output ceiling dependent on a separate long-decode setting.