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

After ~20% of test runs ended early on fallback, rerunning with Antigravity Harness put Opus 5.5 at 58%, Gemini 4 Argon at 54% and GPT-6 Astra at 44%, lifting Argon to second place.

The author argues Google has quietly iterated on the Flash line all year, and a strong Gemini 4 Flash could drive more usage than the expensive flagship Carbon most people won't touch.

A post says Sonnet-5.5 swept all three boards on Apple's new benchmark ahead of Opus-5.5 and Fable-5.1, while Gemini 4 Argon placed sixth of 14.

With Gemini 4 pushing concept generation to single-digit dollars, the author proposes a five-step workflow: spread 40 directions, cull derivatives, pair survivors, push the winner, then lock.

A short reply to theo says something only holds if Gemini 4 Argon writes it, with no further context.

After Pi Harness prematurely killed 20% of runs on refusal paths, rerunning Argon with Antigravity Harness lifted its FrontierCS 2 score from 40% to 54%, beating GPT-6 Astra.

Pi Harness was killing 10 of 48 runs on fallback paths; with Antigravity Harness, Argon scored 54% versus GPT-6 Astra's 44% and Opus 5.5's 58%.

On the latest RSI-Inference Bench, Gemini 4 Argon ranks sixth overall, 5 points behind GPT-6 Astra, with 1.39x inference speedup and the first Day-0 model-support pass outside OpenAI and Anthropic.

A rumor roundup claims Anthropic's IPO week could bring Fable 5.5, GPT-6.1 Terra, Gemini 4 Carbon and Qwen 4, with the poster most excited for Fable 5.5.

An audit found the Pi harness aborted on refusals, cutting ~20% of Argon's FrontierCS 2 runs; rerun in Antigravity, it rose from 40% to 54%, beating GPT-6 Astra and trailing only Opus 5.5.

Gemini-4-Argon enters the RSI-Inference Bench with a 1.39× speedup (4th of 14), 6th overall and 6.7% model-support pass@1, making Google the third provider with a non-zero score.

Finding Gemini understood gritty FPGA details better, the author split work in Antigravity on Google AI Ultra 5x: Flash for planning, Opus 5.5 for coding, Gemini 4 Argon for adversarial review.