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

A leak says Gemini 4 Argon will be announced on September 30 and pushed first to Fairwind network defense testers, is not yet in the model list, and may beat GPT-6 Astra on early benchmarks.

A leak says Gemini 4 Argon rolls out today with Ultra access first, a 1M-token context window and $4 input / $20 output pricing, with early tests placing it in the Opus 5.5 and GPT-6 Astra class.

Google announced Gemini 4 Argon eight days ago but has not shipped it, releasing Nano Banana 2.1 to everyone in the meantime, and says Ultra subscribers and paid API customers will get Argon first.

The author says leaked Gemini 4 output is already on par with GPT-6 Astra and Opus 5.5, with surprisingly strong Arena results, and expects a release today.

A leak says Gemini Cloud will launch a unified agent on October 9 that combines chat, coding and media generation, supports Claude models and has dedicated collaborating sub-agents.

The Argon launch rumor has slid from yesterday to today with no official word from Google; the only new development is AI Studio cutting free caps and raising Pro/Ultra limits, with pricing rumored at $4 in / $20 out.

A reply claims Google plans to release Gemini 4 Flash first, while Gemini 4 Pro will take longer.

A developer testing Gemini Canvas suspects it now routes to Gemini 4 Flash or Flash-Lite, with noticeably worse results, and shares a video of the complaint.

With Google set to cut free-tier Flash access, a user points to daily updates from OpenAI and Anthropic and demands Gemini 4 Argon in AI Studio tonight.

A leak claims Gemini 4 Flash will be available only on Pro plans and Gemini 4 Pro/Argon only on Ultra plans, with a screenshot as evidence.

Google has tightened daily limits for free AI Studio users and raised them significantly for Pro and Ultra, without mentioning Gemini 4 Argon as Argon leaks heat up.

In a reply, a user said they would consider canceling outright if Gemini 4 Pro can't hold up, and that the company should at least add a rate limit.