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

Scores said to have leaked from a Google event show Argon leading Claude Opus 5.5 and GPT-6 Astra on both knowledge work and coding.

A user shares a promo animation generated by the grayscale Gemini 4 agron (gemini 3.8 flash exp-a) and says it seems comparable to Opus 5.5.

Ten days after Gemini 4 Argon launched, the poster says they and other regular users still have no access and are left waiting.

FundaAI's weekly argues Gemini 4 shows scaling still works, with Gemini 4.1's post-training gains in coding and long tasks as the next test, and stays bullish on networking and optical interconnects.

A user says the grayscale Gemini 4 Argon (gemini 3.8 flash exp-a) produced a detailed image, sharing a pelican riding a bike.

Google Antigravity's gray-test gemini 3.8 flash (exp-a) is suspected of actually routing to Gemini 4 Argon, since it even answered the candy-riding test image correctly.

A post shares a benchmark screenshot claiming Gemini 4 Argon beats both Claude Opus 5.5 and GPT-6 Astra, and declares that Google is back.

Google is reportedly testing a new Gemini 4 model called Carbon that already outperforms Claude Opus 5.5 and GPT-6 Astra in internal tests, with a public release possibly coming soon.

A leak says that before Argon is even widely available, a new checkpoint called Carbon is already in internal testing, with coding ability that may rival Claude Opus 5.5.

A post says Google is already testing Gemini 4 Carbon internally and asks whether the unreleased Argon still has any point.

A leak says Google is testing Carbon, the model after Argon, on its coding platform Jetski, with one employee saying it codes like Opus 5.5, while Argon is still not in the model picker.

A user says calling 3.8 Flash in Google's Antigravity routes to a mysterious Gemini 4 model, which they believe is Argon.