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

An Opus 5.5 user publicly asks for a single reason to switch to Gemini 4 Argon, framing the head-to-head rivalry between the two models.

A user questions Artificial Analysis rankings where Gemini 4 Argon no longer matches Astra and falls below the cheaper GPT-6.1 Terra.

Reverse-engineering Antigravity client 2.27.0 turns up only Gemini 4 Argon in High/Medium/Low tiers, no Carbon or 4-Flash codenames, plus BYOM/BYOK support, a Jarvis browser-agent route and several internal codenames.

A Japanese user asked Gemini (Argon, high reasoning) to assess Meiji University's harassment handling; the model found the university defensive and likely to leave victims intimidated and isolated.

The author spent $1 on ZenMux's Halloween blind-box event and won 1M Gemini 4 Argon tokens, explaining the rules and deadline.

On AA's FrontierCode 2.0 engineering-code benchmark, DeepSeek V4.1 Flash 1017 scores 28.5% to lead open models, beating Kimi K3.1 Pro, Gemini 4 Argon and GPT-6 Luna.

Community rumors say Gemini 4 uses element codenames — Pro=Argon, Ultra=Carbon, Flash=Hydrogen — citing a gemini-4-flash-preview string in Google's SDK tokenizer and API screenshots, all unconfirmed.

A user reports Gemini 4 Argon consumed roughly 4M tokens and $10 in Gemini CLI without completing the task, while Claude Opus 5.5 Medium passed all 90 tests using about 240K tokens.

A long post breaks down Artificial Analysis's new Intelligence Index: Opus 5.5 leads, Terra places fifth at 51 and beats Gemini 4 Argon and Fable 5.1 at $0.96 per task.

A leaker says Gemini 4 Flash (Carbon) is dropping soon, with early tests stronger at coding and 3D than GPT-6.1 Terra and possibly beating Argon and Opus as a cheaper, faster option.

A short post asserts there is no scenario where GPT Terra or Gemini Argon beats Fable, without offering evidence.

A developer accepted into Google DeepMind's six-month Antigravity Science Program says Gemini Argon is now at the frontier for biomedical AI and is building new biomed apps with it.