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

A list of five services giving free API keys or credits, claimed to cover DeepSeek V4.1 Flash, Gemini 4 Argon and GPT 6 Astra.

MicroAGI's two robotics VLM benchmarks show Gemini 4 Argon dropping 4.9 points on its headline score at high reasoning effort while costing 4.7x more, with self-hosted nanoannot-1 beating it overall.

The author compares always-on agents like ChatGPT Dot and Grok Bot, citing Google's Argon comeback as proof that culture is harder to fix than capability.

A Japanese user posts a video trying out Gemini 4 Argon, showing the codenamed Gemini 4 model running a task.

A follow-up image post jokes that Gemini Argon is not shaking off the accusations made against it.

Wiz and Google DeepMind upgraded Scan for Good with Argon's reasoning to spot complex bugs and speed fixes, already helping dozens of organizations for free.

On the new MicroAnnotationBench, in-house nanoannot-1 leads at 54.8 while Gemini 4 Argon scores 49.6 at 10x the cost; GPT-6 Astra tops the next-action MicroPlanBench.

Google has opened Gemini 4 Argon inside AI Studio Build to Google AI Pro subscribers, letting users describe an app and have it built, with seven prompts shared.

A joke post answers the question of Gemini Argon's parameter count with garbled characters, highlighting that Google has never disclosed Argon's model size.

A meme image post with exaggerated exclamation marks expressing excitement about Gemini 4 Argon, with no concrete information or testing.

The same prompt was run through Gemini 4 Argon and Opus 5.5 to generate rhino images with no edits, showing their stylistic differences and asking readers to vote.

A Japanese user asked Gemini 4 Argon to compile the question patterns G-test takers found hard, and came away shaken by how tough they were.