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

A roundup of possible AI news this week: Nvidia GTC Berlin, three OpenAI Aurora alpha models in the API, Anthropic testing Fable 5.5, after last week's Gemini 4 Argon.

A user asks whether anyone has tested Gemini 4 Argon on coding tasks and how it compares with Opus 5.5.

A user asks whether Google can compete with Anthropic and OpenAI using only Argon 4, and doubts it can.

After coding in a PHP codebase with Gemini 4 Argon, a developer says model choice should hinge on understanding existing code, precise edits and leaving working code alone, and Argon is a strong alternative for a second approach.

A user combined Gemini 4 Argon with Blender MCP and Unreal MCP to build a horror game scene, saying a fully playable game now feels possible.

A user compares quota across Grok, Opencode, Go Plus and Gemini Pro plans, noting Argon limits on Pro have become reasonable and are counted separately from Opus and G3.8 usage.

The author says Gemini Argon reads more naturally than Claude or GPT, yet Pangram gave a 1,680-word Argon document with two words changed only a 48% human rating.

A user asks whether to pick Gemini 4 Argon or Claude Opus 5.5, framing the two as head-to-head options.

The author says the $8.99 AI Pro plan now lets Antigravity users select Gemini 4 Argon, Claude Opus 5.5 and Sonnet 5.5, plus 4x Gemini usage.

A user says past Gemini mockery shouldn't deter people from Argon 4, claiming it solves problems GPT's latest models can't and pairs well with Opus 5.5.

With Gemini 4 Argon leading Text Arena at 1532, Polymarket traders are betting whether the next Claude Fable can reach 1520+ to pass Opus 5.5 and challenge Argon.

A developer built KillSwitch, a language with fake specs, odd variable naming and injection keywords; the leaderboard says Gemini 4 Argon failed Hello World twice while Opus 5 beat Opus 5.5.