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

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.

The author speculates Gemini 4 Flash will be extremely fast, stay cheap, focus on computer use and agentic tasks, or be scrapped in favor of another Argon version.

On a 100-task benchmark spanning 55 microscope slides and 7 tumor types, Gemini 4 Argon ranks first with 32% pass@1.

ZenMux's Halloween promo lets users who spend $1 in cash PAYG daily open a model blind box; the 16-model pool includes 5M Gemini 4 Argon tokens, with rewards up to $500.

A brief reaction claiming Gemini 4 Flash is genuinely fast, without concrete numbers or test details.

A user reports Gemini 4 Flash generating at roughly 300 tokens per second, warning that an unusually fast model on LMArena is likely it.

A user reports fresh chatter that Gemini 4 Flash carries the codename Hydrogen, and jokes about Google's usual wait before access.

The post compares Argon batch at $1/$5 per 1M tokens with Opus 5.5 standard at $4/$20 and walks through a 50M-in, 10M-out migration cost example.

A user comparing the same prompt and directory says Gemini 4 Argon finds more accurate sources and explains them more clearly than fable-5.1 for chatbot and consulting agent use.

Starting from AI companions on Chinese forums running out of context and "parting ways," the author notes Gemini 4 Argon still has 1M context and hopes Google ships 2M plus memory instead of only chasing coding.

A user says GPT-5.6 and Fable-5.1 failed to hit the point, but handing a summarized problem to Gemini-4-argon produced a breakthrough, questioning whether model progress has plateaued.

A leak claims Google's unannounced model codenamed Hydrogen is likely the Gemini 4 Flash tier, and if it nears Argon at a lower price it could briefly take second place from OpenAI.