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

A leak claims Google is internally testing a Gemini 4 checkpoint called Carbon, which employees use heavily and which reportedly beats Opus 5.5 and GPT-6 Astra in internal tests, possibly launching after Argon.

A leak claims Google employees are already testing a Gemini 4 model codenamed Carbon, even though Gemini 4 Argon has not fully rolled out yet.

A benchmark screenshot claims Gemini 4 Argon ranks first in the latest evaluation, but the specific tests are not named.

A meme video draws Anthropic, OpenAI and xAI with the same panicked face over Gemini 4 Argon's arrival, and its author notes it is a joke, not a verdict.

A post with an image says Gemini 4 Argon, reportedly the codename for Gemini 4 Pro, ranks first in the latest benchmark, without naming the specific test.

Someone extracted Gemini 4 Argon settings from Google Antigravity's program files, showing a separate model entry, thinking-effort controls, and 256K/512K/900K context tiers with 1.3x and 1.8x quota multipliers.

A user says Gemini 4 Argon is already set up in Antigravity and could be released next week.

Argon context window options are showing up in Antigravity while Google prepares to retire older models on November 2, and users complain the migration arrived before the model.

A leak says Google is internally testing a Gemini 4 checkpoint called Carbon that employees use heavily and that reportedly beats Opus 5.5 and GPT-6 Astra, with Argon set to roll out publicly soon.

A leak claims Google is internally testing a new Gemini 4 checkpoint called Carbon, which employees use heavily and which reportedly beats Opus 5.5 and GPT-6 Astra internally, with Gemini 4 Argon also set for public release.

A post jokes that Gemini 4 Argon is the only model being trolled even before it is available to the public.

A poster says that apart from Claude and GPT trading blows, Chinese models have gone silent lately, and the only notable one, Google's Gemini 4 Argon, still has not officially launched, so users should not rush into annual subscriptions.