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

A joke that Anthropic deliberately waited for Gemini 4 Argon to launch before releasing Fable 5.5 to one-up Google.

Google's Gemini 4 launch blog cites Gamow Labs using Argon to analyze newborn genomes, finding a chromosomal rearrangement routine software missed for years, with RareBench at 57% and LabBench at 52.2%.

An Arabic post says Gemini 4 Argon is not just a chat model but also an agent, image model and video model, and is currently being released to Ultra subscribers in Antigravity.

Third-party platform Dot has added Gemini 4 Argon with 1M context and agent workflows, priced 55% below market with USDC and nearly 60% below when credits are funded with $DOT.

The author says Gemini 4 Argon is the closest model to Claude Opus 5.5, which is a notch better on water reflections, architectural detail and frame rate, but Argon costs at least 5x less.

A user found that Gemini 4 Argon says its knowledge runs to March 2026 yet denies the existence of GPT-5, which was released on August 7, 2025.

Tested on misleading API documentation with nonexistent methods mixed in, Gemini 4 caught most fake methods and suggested alternatives, but still confidently accepted one.

The author asked Gemini 4 Argon in the Gemini App to generate a 3D drag racing game and says the output is good and runs without errors.

A user says they give all their tokens to Argon first, calling it the first model to beat Gemini 3 Flash on cost, and notes it is currently 50% off on OpenRouter.

A user reports that Gemini 4 Argon is available to Pro subscribers, but the five-hour usage quota is small and not enough for heavy use.

Google released a safety blog post and full report on Gemini 4, covering the capability evaluations and mitigation work done for the launch.

A developer says Gemini Argon 4 generated five sections for $1.36 but showed no creativity, and that iterating with Haiku is probably better for frontend work.