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

A post claims Google insiders are enthusiastic about Gemini 4 Carbon and say it codes like Opus 5.5, while Argon still has no firm release date.

A poll asks users which upcoming model release they are most excited for, listing Gemini 4 Carbon alongside GPT 6.1 Astra, Grok 4.8 and Fable 5.5.

With Gemini 4 Argon repeatedly delayed, one user jokes that Google should skip Gemini 4 entirely and release "Gemini 5 Delay" instead.

EmirAILab posts a video showing newly leaked Gemini 4 Pro benchmark scores, with the actual figures only visible in the video.

A Chinese user guesses that exp-a on Arena is the rumored Gemini 4 Argon and control may be Gemini 4 Flash, reading it as a sign of a limited test.

A user says they dreamed Gemini 4 Argon would be released on October 18, joking that being a nerd is now infecting their dreams, reflecting community anxiety over the launch date.

The author worries Google will repeat the Gemini 3.5 Pro situation where a model was announced but not released, and calls for Argon and Gemini 4 Flash to launch together or risk losing credibility.

The poster suggests Gemini 4 Argon may be held back for extra safety evaluation because it is too capable, noting Google is already testing the follow-up codenames Carbon and Barium.

Business Insider reports Google is testing a Gemini 4 variant codenamed Carbon on its internal coding platform Jetski, with an employee saying coding feels close to Opus 5.5, alongside Argon and Barium.

A user jokes that Google rejected their request to test Gemini 4 Argon and told them to keep using the Flash models, suggesting Argon is not yet open for access.

A rumor says Google is internally testing a Gemini 4 checkpoint called Carbon that reportedly outperforms Opus 5.5 and GPT-6 Astra and could arrive after Argon.

Business Insider reports that Google staff are testing a Gemini 4 checkpoint codenamed Carbon, with coding said to feel comparable to Opus 5.5, while the announced Argon is still unavailable to regular users.