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

The author made a working realtime thermal camera effect with Argon on the first try, with a video demo attached.

A user says Gemini 4 Argon produces high-quality output and follows instructions well, and that they have not yet hit the 5-hour usage limit, suggesting Google may have something special.

The author says Argon is a natively multimodal sparse MoE model that can output up to 1M tokens in one response, and compares its strengths against the shortcomings of Opus 5.5, GPT-6 Astra and Grok 4.7.

The poster says Google's Argon 4 and Meta's Wonder Full are good models but don't come close to Opus 5.5, leaving Anthropic on top.

The poster lists three new models released over the past month, including Gemini 4 Argon, and still calls Opus 5.5 the best model by a mile.

A Japanese illustrator used Gemini 4 Argon to generate a twintail look for their original character Kagari Kiryu, saying the result came out better than expected.

Wiz's AI cybersecurity arena added Codex and Antigravity harnesses, with Gemini 4 Argon first at 77.2% success and GPT-6 Luna the best value at 64.5% for $0.05.

A short post argues that Muse performs poorly and Gemini 4 Argon falls far short of Opus 5.5, and asks whether others agree.

A post says that just 20 hours after Gemini 4 Argon launched, people had already built eight wild projects, and lists them.

A user complains that Droid subscribers have more access to Gemini 4 Argon than Gemini Pro subscribers, and get it at 50% off.

A roundup notes a month of dense releases from Claude, ChatGPT and now Gemini 4 Argon, and expects Chinese models to lead the next wave.

Of 59 X posts about Gemini 4 Argon one day after launch, 18 were positive, 19 neutral and 18 negative; the most common praise was its Japanese writing, and the most common complaint was the 5-hour quota running out quickly.