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

A user reads Google's codenames as signals: Argon (an inert gas) for stable, defensive security and the rumored Carbon for a breakthrough in coding fundamentals, and wonders if Gemini 4 Argon will only ship after Carbon launches.

The author predicts Gemini 4 Pro will fail badly, saying Google's AI team and leadership have dropped the ball, especially in the browser, and that Google won't recover.

A Chinese post relays that Gemini 4 Carbon reportedly matches Opus 5.5 at coding, and jokes that claiming to be the best while never releasing means no one can test it.

The poster says a DeepMind employee publicly mentioned RSI and that Google is reportedly testing a new checkpoint, Gemini 4 Carbon, which is stronger than Argon, though Argon still hasn't shipped.

A Chinese post mocks Google's delayed Gemini 4 Argon, saying that by the time it launches, Opus 6.0, GPT-7 Bel, Grok 5.0 and others will already have shipped and been iterated on.

Gemini's web app replaced extend thinking with low, medium and high effort options, and Logan says Gemini 4 Argon is coming soon, echoing what he said before Gemini 3.5 Pro.

The post says it is the weekend and no one except MiniMax releases new models on weekends, so Gemini 4 Argon will not arrive in the next two days.

Business Insider reports Google employees are testing an internal Gemini 4 checkpoint called Carbon, with one rating its coding near Claude Opus 5.5, while the public Argon was picked from Barium-B.

A short post with an image says the author can't wait for Gemini 4 Argon and hints that competitors may be nervous.

A reply speculates that Google is deliberately weakening current models so users will move to the unreleased Gemini 4 Argon.

A developer complains that Gemini 3.8 took more than four hours to fix one task, far worse than before, and urges Google to release Gemini 4 Argon.

The October 10 AI daily brief covers five items, including a rumor that Google will release Gemini 4 Argon and is testing Carbon internally, plus Codex, Qwen-Image, and a Microsoft decision model.