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

A designer built a web tool with Gemini 4 Argon, Antigravity and AI Studio for drawing swashes on existing typefaces, currently sourced from Google Fonts.

Google opened Gemini 4 Argon and two Claude models to AI Pro users on Oct 19, but users report quotas draining fast and Argon working only in the app, with no official word on weekly quota math.

Switching only the harness to Antigravity lifted Gemini 4 Argon's FrontierCS 2 score by 14 points, which the author credits to graceful fallback recovery keeping the loop alive until tasks finish.

Auditing traces showed Pi Harness ended ~20% of runs at fallbacks such as security tasks; rerunning Gemini 4 Argon with Antigravity Harness put it ahead of Astra and just behind Opus 5.5.

The author spent ~$5k running agent fleets on H200s for GPU kernels and found human-guided fleets progress over 2x faster than Astra-guided ones, while Gemini Argon kept stopping for security reasons.

The author argues better AI text comes from using well-tuned models rather than prompting weak ones, citing Claude Opus 5.5, Haiku 5.5 and Gemini 3.8 Flash / 4 Argon.

A user reports Gemini Argon confused Orville and Wilbur Wright's flight records and had to be corrected, arguing hallucinations are not solved across model families.

An October 26 AI news roundup notes Google's October 23 showcase of Gemini 4 Argon building 3D game worlds and data visualizations.

Replying to another post, the user asks how Gemini 4 Argon would have performed on that image identification task, saying Gemini models often punch above their weight there.

The author configures a Google AI Studio API key in DeepSeek Harness and enters the model ID gemini-4-argon to use it, with usage first drawing on the $10 monthly Google AI Pro credits.

A two-model workflow: Gemini 4 Argon quickly builds core logic and interactive prototypes at high speed and low cost, then Claude Opus 5.5 handles the 1-to-10 optimization.

Google is reportedly testing a new Gemini 4 model called Carbon, with early testers citing gains in 3D generation, spatial reasoning and output quality over Argon, plus coding compared to Claude Opus 5.5.