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

A video shows Gemini 4 Argon building a Minecraft clone in 54 minutes, which the author calls the best model ever.

A video showcases 10 wild examples of Gemini 4 Argon, claiming it makes other models look like nothing.

A developer who stress-tested both models on real projects for three weeks lists four areas where Gemini 4 Argon wins and four where Claude Opus 5.5 does.

On the same 100 hard questions, Gemini 4 Argon got only 50 right but made up just 7 answers and said "I don't know" 39 times, trading accuracy for honesty versus Claude Fable 5.1 and GPT-6 Astra.

A creator shares a Gemini 4 Argon / Antigravity prompt for a 45-second motion graphics video, with code and prompts on GitHub.

On the same Eiffel Tower build, the post claims Opus 5.5 took 2 hours and $33.45 while Gemini 4 Argon finished in one shot for $0.23, roughly 145x cheaper.

A user jokes that Gemini 4 Argon's output quality swings wildly between tasks, sometimes near Luna/Haiku, sometimes beating Grok, Sol and Sonnet, and close to Opus 5.5 on an AWP skin test while cheaper and faster.

Same Eiffel Tower build task cost $33.45 on Opus 5.5 versus $0.23 on Gemini 4 Argon, a 145x gap, with both outputs shown.

Comparing $200 plans, the author says Claude Max users can't use up Opus 5.5 while Gemini Ultra users must ration Argon usage.

A user claims Argon is only available in web Gemini and not in Antigravity, calling it useless and questioning the late reporting.

A comparison chart claims Gemini 4 Argon has a 15% hallucination rate, below GPT-4.1 Nano's 18%, while GPT-6 Fable and Claude Fable 5.1 sit at 92%–95%.

The author asks whether everyone hyped Gemini Argon then went back to Claude, noting little chatter from Argon users now.