Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

A poster asks whether anyone would pay $200 a month for Gemini 4 Argon if Claude Opus 5.5 still performs better.

A streamer posts that a Gemini 4 Argon stream will start in 30 minutes.

A video compares Gemini 4 Argon and Opus 5.5 on the same prompt, and the author says Argon made Opus 5.5 look bad.

A user reports that Argon is now available in the Gemini app, can read folders directly to audit or write code, and has limits separate from Antigravity, with the 5-hour cap burning fast but a generous weekly quota.

A thread rounds up Argon's launch notes, pricing and model card: $2/$10 per million tokens until November 30, then $4/$20, a 95% discount on cached input, and a focus on long agent runs and prompt-injection resistance.

The poster says a friend who knows no coding built three working web apps in 30 minutes using Gemini 4 Argon, and asks whether developers are in trouble or this is the future.

A Japanese developer says yomiyasu v2.0.0 is finally taking shape, but Gemini 4 Argon responds so slowly that development takes much longer, so they are optimizing orchestration to ship as soon as possible.

A clickable drawer-cabinet diagram uses the sparse MoE noted in Gemini 4 Argon's official model card to show how a router activates only some experts per token.

The Japanese AI podcast previews an episode on AI self-replication risk and takes a closer look at the now-public Google Gemini 4 Argon and Meta Muse Wonder.

A developer reports that in a single long chat, Argon's fourth prompt re-bills 140k+ accumulated input tokens, and that terminal logs can break the cache.

After Gemini 4 Argon's general availability, two leaderboards disagree sharply, ranking it first on the Vals Index and around ninth on Artificial Analysis, and the poster asks which to trust.

A Korean user says it took seven attempts to get an image out of Gemini Argon, and the result still differs noticeably from the original in detail.