Google has launched Gemini 4 Argon, a frontier model for complex software engineering and enterprise workflows, Sundar Pichai said on X. The model is available globally in the paid tier of the Gemini API and the Gemini platform, and for Ultra subscribers in Google Antigravity S2.
According to the company, Gemini 4 Argon handles multi-turn investigations, multimodal document analysis and scientific reasoning. The model supports a combined input-output token context window of up to 1M and was trained using reinforcement learning with multi-step reasoning data S8. Its knowledge cutoff is March 2026, the model card states.
Automatica reported in July that Google had previewed a similar model for internal use. Gemini 4 Argon was evaluated on agentic coding, ML engineering, science, math, multimodal, long-context and cybersecurity benchmarks, the model card says. It also lists known limitations including occasional hallucinations, slowness and timeout issues S8. No independent test results are included in the announcement.
A third-party evaluation from Bindu Reddy placed Gemini 4 Argon behind Fable and Astra but ahead of open-source models on LiveBench 2.0, a benchmark designed to reflect real-world usage S3.