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

Testing tricky PDF-to-Markdown conversion, a user finds Gemini 4 Argon worse at OCR than Gemini 3.0 Flash, saying it prefers to fabricate content rather than extract messy text as-is.

After reading papers, the user rates Opus 5.5's writing taste, Gemini 4's search strength, and Astra's critical but idiosyncratic lens.

A user trying T3 Code found it only listed Flash and stale cached models instead of Gemini 4 Argon; reinstalling the ACP runtime fixed detection, and they may switch to it full-time.

A report says Gemini 4 Flash or another Flash model is being tested with select users, described as very fast with Argon-level power for daily use.

A Turkish post lists new releases including GPT-6 and Gemini 4 Argon, notes small fast models now cost around $0.10-0.15 input, and advises sticking with one model for 30 days.

A demo shows Gemini 4 Argon editing a gaming video from a single prompt, adding SFX and music, zooms and effects, cuts, captions and reframing.

The author used Gemini 4 Argon High in Antigravity to generate an interactive Three.js r184 scene turning three deep-sea cnidarians into 19th-century brass clockwork automata, with multiple plate environments and anatomical callouts.

A French user shares a grading result where Gemini 4 Argon hands out a too-generous 4/20 score.

The author says Gemini 4 Argon is unmatched at copy, literary writing and translation, yet hallucinates and repeats mistakes from two years ago, making it too inconsistent to rely on.

Rebuilding the Chrome dino with ragdoll physics from one prompt: Opus 5.5 took 38m55s for $8.53, Kimi K3.1 19m18s for $1.25, and Gemini 4 Argon 4m34s for $0.19; the author ranked Opus best but noted Argon was 45x cheaper.

A developer says despite all the Gemini 4 Argon hype he has never opened it, because Claude never gave him a reason to switch.

A from-scratch Three.js animation benchmark with no external assets ranks Gemini 4 Argon second, tied in the High tier with Sonnet 5.5.