Google has unveiled Gemini 4 Argon, a frontier model designed for complex, long-running work in software engineering, finance, law and cybersecurity.[1] It is initially being released to selected cyber defenders through Google’s Fairwind programme while the company strengthens safeguards before wider availability.[1]
On Google’s disclosed comparison, Argon led outright on 12 of 18 benchmarks and tied for first on another, giving it more top results than GPT-6 Astra or Claude Opus 5.5 within that test set.[2] Its results included 77.9% on the long-horizon DeepSWE coding benchmark, 51.3% on Zapier’s AutomationBench and 91.7% on a long-video understanding test.[1][2]
However, it did not sweep the field: GPT-6 Astra retained leads in some software and scientific-terminal tasks, while Claude Opus 5.5 led in terminal-agent and post-training evaluations.[2]
Independent testing presents the race as even closer.[3] Artificial Analysis scored Argon at 53 on its Intelligence Index, one point behind Claude Opus 5.5 at comparable high-effort settings.[3] Argon led that comparison on automation, science coding and several knowledge measures, while Claude used substantially fewer output and reasoning tokens per task.[3]
That makes Argon a frontier competitor rather than evidence that Google has decisively overtaken every rival.
Google is also offering a distinctive technical proposition.[1] Argon has an output-token limit of one million, intended to sustain unusually long reasoning and agentic workflows.[1] Google says its engineers are already using the model for debugging, codebase migrations and algorithm design, including work to migrate large C and C++ codebases to Rust.[1] These are company-reported examples, not yet independently reproduced results.
Cost could strengthen Google’s position once access expands.[1][2] The introductory API price is $2 per million input tokens and $10 per million output tokens, later rising to $4 and $20 respectively.[1] Broader access is expected to begin with paid API customers and Google AI Ultra subscribers, but Google has not given a firm public release date.[1][2]
The caution is that performance claims still need testing in normal production conditions.[2][4] Bloomberg reported internal scepticism at Google about Argon’s performance in areas including coding, despite its strong benchmark scores.[4] With access currently restricted, outside developers cannot yet settle the difference between a strong launch table and reliable daily performance.[1][2][4]
Argon therefore gets Google back into the ring, but not across the finish line.[1][2][3] It brings competitive benchmark breadth, long-running capability and pricing pressure to a market led by OpenAI and Anthropic.[1][2][3] Whether it changes the order at the frontier will depend on what happens when customers can finally use it at scale.
Supporting sources
[2] VentureBeat: Google unveils Gemini 4 Argon
[3] Artificial Analysis: Gemini 4 Argon versus Claude Opus 5.5
[4] Bloomberg: Google faces internal Gemini 4 scepticism
Source: Google