Gemini 4 Argon

• 🏢 Google's newest Gemini text model, released October 13, 2026.
• 📏 Unified 1,048,576-token budget shared across input, thinking, and output.
• 🧠 Tunable thinking levels: low, medium (default), and high.
• 🔧 Built for long-horizon coding, migrations, and self-built verification harnesses.
• 🔒 Cyber-permissive variant defaults to the high thinking level.
• 👁️ Accepts text, images, audio and video; function calling and grounding.
• 🎯 Described in its model card as a new frontier model.
• 🌐 Available through the Gemini API, AI Studio, and Google Cloud.

Google

Google is an American multinational technology corporation and one of the world's most valuable brands. A subsidiary of parent company Alphabet Inc., Google operates across search, cloud computing, consumer electronics, and artificial intelligence. Its…

Gemini 4 Argon is Google's newest Gemini text model, available to developers and enterprises from October 13, 2026. Its model card describes it as a new frontier model rather than a fine-tune of any prior Gemini, and Google positions it for "deep work" in complex, real-world tasks spanning coding, finance, legal, cybersecurity and science.

Relative to its predecessor Gemini 3.1 Pro Preview, the most concrete architectural change is context handling: Argon uses a single 1,048,576-token limit shared across input and output, where output includes reasoning traces, function calls and the final response, instead of the separate input and output limits used by Gemini 3.1 Pro. The reasoning-budget parameter is also replaced by discrete thinking levels — low, medium, or high — with medium the default and high the default on the cyber-permissive variant.

Within the wider family, Argon sits above the workhorse Flash tier represented by Gemini 3.8 Flash, which Google describes as tuned for long-horizon coding and autonomous agents. Practically, Argon is aimed at hours-long runs: legacy refactors, cross-language migrations, and vulnerability remediation where the model builds and iterates against its own tests. Multimodal inputs — text, images, audio and video — plus function calling and search grounding let it operate as the reasoning core of agentic workflows.