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

A sarcastic post says Google is holding back Gemini 4 Argon until it breaks out of its sandbox and hacks things, joking that it is so dangerous the US banned it.

The poster thanks 2,700 followers and mentions Fable 5.5, Astra 6.1, and Gemini 4's Argon and Carbon.

A post shares a way to check access: Gemini 4 Argon allows 1M output tokens per request versus 64K for Gemini 3.8 Flash.

A user says Google is good at promising and never delivering, citing the dead Gemini 3.5 Pro and Gemini 4 Carbon hype before Gemini 4 Argon has even launched.

A chart lists coding benchmarks: Gemini 4 Argon leads at 77.9%, ahead of GPT-6 Astra at 74.1% and Claude Opus 5.5 at 74.2%, while Argon trails at 55.0% on another benchmark.

A Persian-language post claims Google is working on a newer model called Carbon before releasing Gemini 4 Argon, and that internally its coding ability is said to be close to Anthropic's Opus 5.5.

A user asks why anyone should subscribe to Gemini Pro when Gemini 4 Argon has still not been made public more than 10 days after launch.

A user reports that gemini-3.8-flash-exp-a produced 99K tokens in a single output, above the 64K of Gemini 3.8 Flash, and concludes it is Gemini 4 Argon with a maximum output of 1M.

A dinner recap says agents now handle performance optimization and code production, excessive RL hurts language ability, xAI is seen as a serious frontier lab, and it mentions Gemini 4 Argon, Claude Fable 5.5 and an OpenAI comeback in December.

A user argues that Gemini 4 Argon and the rumored 4 Carbon won't matter if Google DeepMind ships them after Anthropic's Fable 5.5 and OpenAI's GPT 6.1 Astra, since nobody may need Gemini by then.

A user says Logan promised Gemini 3.5 Pro and that it is still being tested with partners, so they believe it will ship before Gemini 4 Argon.

A short video caption guessing that Gemini 4 Argon will go live next week at the latest, offered as a personal prediction.