Gemini 4

Gemini 4 Argon is fully available to paid users, but the Anti-Gravity feature is limited to Ultra subscribers, and Pro users cannot use it.

Gemini 4 Argon's maximum output has been increased from 64,000 tokens to 1 million (referring to the output limit, not the context window)

Google has decided to officially abandon the checkpoint versions of Gemini 3.5 Pro.

Google published a new member of the Gemini series, Gemini 4 Argon, on its official blog.

Models such as DeepSeek are priced much lower than Gemini 4.

The video creator estimates Gemini 4 is about 70% cheaper than Opus 5.5.

According to the Agent Arena data cited in the video, Gemini 4 ranks eighth on that leaderboard.

Gemini 4 ranks slightly above Sol 5-6, but below the latest models such as 6 Sol, Fable 5 and Opus 5.

In the "island shelter" 3D generation test, Gemini 4 showed problems such as trees falling in wrong directions and misplaced elements; the author considers this performance "certainly not the best", while Sonnet 5.5 was rich in detail, coherent and more visually appealing.

In the "Tower of Babel" 3D generation test, Gemini 4 showed floating figures, characters stuck motionless, rigid arrangement and problems with water handling, while the figures in GPT 6.1 Sol had clearer design intent and behavioral intent, with overall completion a level higher.

Gemini 4 enters the Pareto frontier, but sits in the middle range.

The video creator estimates Gemini 4's cost at about two-thirds that of 6 Sol.

In the 3D generation test with the "Golden Gate Bridge" prompt, Gemini 4's coherence and 3D texture were good, the water texture was impressive, and the UI was thoughtfully designed; the author considers it on par with or even slightly better than Sonnet 5.5, but GPT 6.1 Sol is superior.

In the game generation test, the mini-game control system generated by Gemini 4 was "pretty bad", making the gameplay hard to understand, and characters got stuck; by contrast, the game generated by GPT 6.1 Sol was at least playable and characters did not run around everywhere.

At the end of the test, the uploader listed the limitations: the difficulty order of the five questions was his own arrangement and not a standard evaluation; each question was run only once, with no repeated sampling; the prompts were written as casual one-off questions, so what was tested was "the level of a casual one-off question," not the model's upper limit.

In the test, all five questions were sent at once, with no follow-ups, no hints, no changes to the questions, and networking was prohibited.

In the coding question (writing a Breakout game in HTML, with external resources prohibited), both Anti-Gravity and Vertex AI produced complete, runnable code in one go.

Google Gemini 4 generated a copy-paste-style small scene in this test, with a scale that does not meet the requirements of a "hyperscale cluster".

The uploader's conclusion is that Gemini 4 wins on this problem, but emphasizes that testing only one problem cannot decide the winner and that results may differ with a different problem; DeepSeek called the official API directly and did not specially configure the environment, which differs from Gemini's environment, so speeds are not compared.

Gemini 4 was closer to a runnable state on its first generation, whereas DeepSeek V4 Pro's code, though longer, made elementary mistakes in basic logic.