Business

Google announces Gemini 4 Argon AI model, but you can’t use it yet

Google is stepping back into the high stakes race for frontier artificial intelligence with the announcement of Gemini 4 Argon. After spending much of the summer focusing on smaller, faster Flash models instead of the promised Gemini 3.5 Pro, the tech giant is now pivoting toward raw power. According to the company, Argon delivers industry leading performance specifically tailored for coding, complex knowledge work, and cybersecurity. However, there is a catch for the general public: despite the fanfare, the model remains locked away from consumer access for the time being.

While external users wait, Google reveals that its own internal engineers have already put Argon to work on massive infrastructure projects. The company reports that by utilizing fleet wide telemetry data, Argon helped shave 300 terabytes of memory usage across its global data centers. Even more impressive is how the AI has handled legacy code; Argon agents have been migrating vast quantities of C and C plus plus code over to Rust, including hundreds of thousands of lines within the Fuchsia OS Zircon kernel and critical core libraries like re2 and libgav1.

To prove these aren’t just empty claims, Google released several benchmarks showing Argon outperforming major rivals. On the DeepSWE software engineering test, it scored nearly seventy eight percent, edging out competitors like GPT 6 Astra and Opus 5.5. The company also highlighted strong results in economic analysis via the Vals Index test, suggesting a level of reasoning capable of handling long horizon tasks that previously stumped AI models.

Even though a public release date hasn’t been set, Google is already laying out the financial details for developers through upcoming API pricing. Input tokens will cost two dollars per million while output tokens will run ten dollars per million, with significant discounts available for cached inputs. Perhaps most notably, Argon will feature a staggering one million token output limit—a massive leap from the sixty four thousand token ceiling seen in previous versions—which Google says will allow users to tackle far more daunting projects in a single pass.

Comments are closed.