Google announced Gemini 4 Argon on September 30, 2026, with access initially limited to vetted security defenders through the Fairwind Program and early testers such as Wiz. It reached developers on October 13 through the paid tier of the Gemini API and AI Studio, with Antigravity available to Google AI Ultra subscribers. The model ID is gemini-4-argon, and a separate gemini-4-argon-cyber-permissive variant is aimed at authorized defenders. Argon is a sparse mixture-of-experts transformer with a March 2026 cutoff. It accepts text, code, images, audio, video, and PDFs, but outputs only text and code.
The most notable change is the output window. Older Gemini models handled roughly a million tokens of input and returned at most 64K tokens. Argon raises the output limit to 1M tokens, using a unified window of 1,048,576 tokens shared across input, thinking, and output. A 100K-token prompt therefore leaves room for nearly 950K tokens of answer. The default output cap remains 65,536 tokens, and developers must raise max_output_tokens and use long decode to go further. Long decode can run for several minutes, so it suits background jobs.
The model also offers three thinking levels: LOW, MEDIUM (the default), and HIGH. The benchmark picture is uneven. Argon leads on repository-level software engineering but trails on terminal work, and a careless API call can cost $10 to $20. Those weaknesses matter more to daily users than the headline averages.
For developers, the practical takeaway is that Argon removes the need to chunk large generations and re-feed context on every turn. Its strengths are clearest in long, complex coding tasks, while its weaker terminal performance and higher per-call costs call for careful budgeting and guardrails in scripts.
Source: Towards Data Science · Summarized by HeadlinesBriefing