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

An Indonesian-language user thanks Gemini 4 Argon for downloading restricted Drive files and producing an AI version audit.

The poster advises running both Gemini Argon and GPT Astra on the medium setting to avoid wasting usage quota.

Gemini 4 Argon reportedly places sixth on Apple's RSI-Inference Bench and is the first non-Anthropic/non-OpenAI model to pass a Day-0 Model Support task.

On the new FrontierCS 2 benchmark, where agents fix real CS research code without seeing the paper, Opus 5.5 passes 58%, GPT-6 Astra 44% and Gemini 4 Argon 40%, with Argon nearly catching Opus under a different harness.

Jensen Huang dismisses neocloud financing worries, pointing to strong demand across OpenAI, Anthropic, Meta, Grok and the just-released Gemini 4 as AI hits an inflection point.

With Antigravity, Gemini 4 Argon roughly ties GPT-6-Astra on the RSI-Inference Bench and is the first non-Anthropic, non-OpenAI model to pass a Day-0 Model Support task.

The poster says their eight accounts were untouched, noting Jio-channel Google AI Pro users can use Argon in Antigravity while student-channel and CPA proxy users were purged.

A user says they will now use Gemini 4 Argon extensively inside Antigravity, with a screenshot attached.

On the same 48 tasks, switching from the Pi harness to Antigravity, which retries content-filter refusals, flipped 8 tasks to pass and lifted Gemini 4 Argon's pass rate from 40% to 54%, second behind Opus 5.5.

A brief take that Google's Antigravity coding environment paired with Gemini 4 Argon makes a powerful combination, without specific test details.

The author ranks models 34 days on, saying Opus 5.5 still leads, with Claude Haiku 5.5 and Gemini 4 Argon next, followed by Kimi K3.1 Pro, Meta Muse and GPT 6.1 variants.

Gemini 4 Argon joins RSI-Inference Bench at #6 of 14 overall, just behind GPT-6 Astra, with 81% hidden-test pass rate and 1.39x speed-up for 4th in inference optimization.