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

A leak says Gemini 4 Argon has appeared in a Gemini API documentation entry with a 1M token output limit and $2/$10 pricing, and the community expects it to be callable next week.

A video claims Gemini 4 Argon has entered the Agent Arena top 10 for the first time, up from Gemini's previous best of #22 with 3.8 Flash.

The author gave Gemini 4 Pro a screen recording and asked it to recreate Duolingo, as a hands-on demo of its front-end reproduction ability.

A post says Gemini 4 Argon is already in the API docs with a public launch imminent, landing the same week as GPT-6.1 and Anthropic's Fable 5.5.

A joke that Google is putting compute resources into space projects, which is why the Argon-codenamed Gemini 4 Pro has yet to ship.

The author says that feeding Gemini 4 Pro a screen recording of an app like Duolingo lets it reproduce the interface design and functionality based on what it sees.

Paid users of Google's Antigravity IDE can use Opus 5.5 and Sonnet 5.5 directly, and the team also teased a broader rollout of Gemini 4 Argon.

User Inmerson posted a video titled "Birds" labeled GEMINI 4 PRO, showing the model's generation output.

A Korean post says Gemini 4 Argon has beaten Opus 4.8 on Agent Arena and jokes that it should not be competing only at the Haiku level.

The poster says Gemini 4 Argon now beats Qween 3.8 and Opus 5 on the Web Dev Arena, with a ranking screenshot attached.

A Korean user looks forward to seeing Fable 5.5, GPT 6.1 Astra and Gemini 4 Pro next week, and jokes about whether Gemini 4 Pro will suffer another honorable death.

User SadAlbert10 says the claim that Gemini 3.1 is routed to Gemini 4 Argon is fake, and complains that Gemini still searches the web aggressively even when told not to.