I was wrong. If you walk in the rain often, how can your shoes not get wet? I had just grumbled about the Gemini team’s snail-like pace, and then they turned around and brought out Gemini 4 Argon to slap me in the face—too swift, completely without warning.
Judging from the published benchmarks, the numbers this time really are frightening enough.
The single-output limit has been raised directly to an industry-rare 1 million tokens, with pricing at $2 per million input tokens and $10 for output.
I should specifically mention: currently, the various million-token contexts on the market mainly refer to "input"; single-output limits are usually very conservative. A single output of one million tokens can handle more complex continuous tasks.
Argon’s pricing is clearly very targeted: the same price as GPT-6.1 Sol, with performance surpassing GPT-6 Astra. The same price as Claude Sonnet 5.5, with performance surpassing Claude Fable 5.1.
The benchmark data feels extremely, extremely, extremely unreal, directly grinding Astra and Fable into the ground.
In the official published battle report, Gemini 4 Argon can autonomously convert Fuchsia’s 800,000 lines of kernel code into Rust, can also optimize memory for an entire data center to squeeze out over a hundred TiB of space, and even took first place in cyber offense-and-defense evaluations.
But currently, only trusted defenders under the Fairwind program can use it. The official line is that it is actively cooperating with relevant departments for pre-release voluntary review, needs to continue collecting feedback and strengthening security protections, and only later will it gradually open to paying users.
I’ve always thought that Google is the one that doesn’t say it out loud but uses practical actions to slow things down.
The official blog post is here: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/