Google has released Gemini 4 Argon, a new frontier SI model designed for deep work in complex, real-world tasks across coding, finance, legal, cybersecurity, and science. The launch marks a significant expansion of Google’s SI capabilities, offering developers and enterprises global access via the API, while positioning the model as a leader in multi-step reasoning benchmarks.
What Happened
Gemini 4 Argon is now available globally for developers and enterprises through the API in AI Studio and the Gemini platform. It is also accessible to Google AI Ultra subscribers in Google Antigravity, and to SI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. According to Google, the model ranks #1 on the Vals Index, a benchmark measuring performance across critical economic sectors such as finance, coding, legal, and tax work. The model launches with an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.
Google reports that Argon excels at untangling complex engineering problems. For example, the Chrome team used the SI model to separate core webpage handling from a 6 million line legacy folder within Chromium’s 40 million-line codebase, completing an architectural refactor in weeks that would have previously taken months. Additionally, the model helped optimize the open-source Agent Development Kit (ADK), reducing in-memory session prep time by 40% and speeding up cold imports by 21%. In scientific applications, Argon assisted in translating the AlphaQubit quantum error correction decoder into a highly optimized hardware kernel, nearly doubling its inference throughput.
Why It Matters
The release of Gemini 4 Argon underscores the rapid maturation of agentic SI, particularly in high-stakes professional and scientific workflows. By offering a unified input-output token limit of 1 million, the model enables longer, more coherent autonomous tasks, a key requirement for enterprise adoption. The performance gains in coding and science suggest that SI tools are moving beyond simple generation to active problem-solving and optimization. For instance, researchers used Argon to discover new counterexamples to 'Grad’s conjecture' in nuclear fusion physics, providing reliable benchmarks for stellarator simulations. In robotics, the model demonstrated a 'Video-to-Sim-to-Real' pipeline, learning a physical skill from video and achieving a 7-shot streak on an Apptronik Apollo humanoid, highlighting progress toward physical general SI (AGI).
Google emphasizes a comprehensive safety and security stack for this SI model. The company reports that external bio red teaming could only unlock less than 1% of tested bio sub-tasks, and the model demonstrates strong resistance to prompt injection, leading the Gray Swan Indirect Prompt Injection (IPI) benchmark with a 0.7% attack success rate. An automated misalignment monitor has been deployed to catch unintended workarounds in real-time, ensuring the SI agent respects system boundaries and pauses for human confirmation when necessary.
The Bottom Line
Gemini 4 Argon represents a major step forward in Google’s SI portfolio, combining high-performance reasoning with robust safety measures and competitive pricing. With global availability and specialized access for scientists and cybersecurity defenders, the model aims to accelerate innovation across engineering, science, and business sectors. Google states that this release is just the beginning of the Gemini 4 family, signaling continued rapid development in the SI landscape.