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

Issue 001 of a Serbian weekly tech magazine lists Gemini 4 Argon alongside OpenAI DevDay and Starship's first orbital flight as the week's top stories.

A Chinese post says Gemini 4 Argon places just 8th on the Code Arena: WebDev leaderboard, behind GPT-6-SOL, which the author says doesn't match Google's promotion.

A leak claims Opus 5.5 and Sonnet 5.5 are coming to Antigravity's Pro and Ultra plans, with Gemini 4 Argon possibly joining soon, making Google's Ultra plan more worthwhile.

A user complains that Claude Sonnet 5.5 has already appeared in Antigravity while they are still waiting for Gemini 4 Argon to arrive.

A post asks developers which model offers the best balance of speed, accuracy and code quality among Opus 5.5, Sonnet 5.5, GPT-6 Astra and Gemini 4 Argon.

The post says gemini-4-argon-high leads the LMSYS text arena at 1525 Elo, and contrasts flagship models at $15–25 per million tokens with lighter architectures at $0.5.

Arena data from October 2 shows Gemini 4 Argon High first in the English text arena at 1533, but only tenth across 10,030 Agent Mode sessions at a median cost of about $0.67 per task.

The poster argues Argon's issue is not benchmarks but timing, since Fable 5.5 is already in progress and 6.1 Astra may follow, and urges shipping Argon now.

A poster says Gemini 3.8 Flash has slowed and Plus users lost Pro model access as compute moves to Gemini 4 Argon, which they expect to stay off subscription tiers.

A video compares Gemini 4 Argon and Opus 5.5 on the pelican-on-a-bike prompt; the author says Gemini did well for the first time, but Opus 5.5 was more polished in lighting and color.

Free users keep only Flash-Lite, AI Plus is limited to Flash-Lite and Flash, and only AI Pro/Ultra keep Pro models, with usage now metered by compute, refreshed every five hours and capped weekly.

A post relays that a DeepMind researcher with free access to Gemini 4 Argon still switches to Gemini 3.8 Flash, since speed matters more than raw power for straightforward tasks.