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

A developer ships a 600M-parameter Persian vision model built on Nemotron 3.5 Parse, claiming it competes with Gemini Argon on their benchmarks.

GBENCH posts GPT-6.1 Terra coding results placing it between Gemini 4 Argon and Claude Opus 5.5, and discusses coding-assistant usability differences.

Gemini 4 Flash is reportedly reaching ~300 tokens/sec on LMArena, which the author says would be a major edge if it delivers Argon-level intelligence at Flash pricing.

Opus 5.5, Sol 6.1 and Gemini 4 Argon all sit on the Pareto frontier, but Opus is more read-efficient while Argon burns more output tokens to reach the same score.

FrontierCode 2.0 separates model tiers more sharply: Muse Wonder 1.0 and Gemini 4 Argon top DeepSWE v1.1 but land mid-pack on FrontierCode 2.0.

In a multiplayer autoresearch run, Gemini 4 Argon set each frontier and rose from 0.625 to 0.938 in round 3 by citing earlier failed attempts; Fable and Sol then built on its strategy.

The poster claims the insanely fast Gemini 4 Argon seen in Arena testing is actually the "hydrogen bomb" Gemini 4 Flash at 300tps.

A leak claims Gemini 4 Flash runs around 300 tps, faster than GPT-6.1 Terra, with Argon-level intelligence and pricing below Argon, near Sonnet 5.5.

Citing another user, the post says Gemini 4 Flash reaches 300 tokens per second on LMArena and asks how many would switch if quality holds.

A leak says Gemini 4 Flash hits close to 300 tokens per second; if it matches Argon-level performance at Flash pricing it would be hard to beat, and the poster plans to test it on Arena.

According to Lyra, the fast Gemini 4 on Arena is the Flash variant at around 300 tps; it isn't released yet but may hit Arena soon.

The poster notes DeepSeek runs around 400 tokens/s and MiniMax M3.1 Flash near 1000 tps, so the real question is whether Gemini 4 Flash can outperform both on actual quality.