Gemini 4 Argon is out. The best demo has two camera angles.

What actually launched

Argon accepts text, images, audio and video; its model output is text, including code. The striking visual results circulating around the launch often come from software it controls or code it writes. Keeping those roles clear tells you which parts of a result may remain editable.

Google’s launch examples combine Argon with Antigravity, Blender, Three.js and other generative models. A camera, a mesh and a script offer different kinds of control from a finished clip. They also leave different kinds of work to the person finishing the project.

Sources: Google [1] · Google DeepMind [3] · Google [2] · Google / YouTube [14]

Two angles. One room. A useful production question.

Google DeepMind creative producer George Kenwright describes a pipeline built from two green-screen stills: Argon constructs a room in Blender and recreates the viewpoints; Nano Banana supplies its visual treatment; the original person is composited back in. His comparison image shows the source plates, an untextured room and the finished backgrounds.

The attraction is shared geometry. If a window moves, both views have a common scene to return to. That could make revision less dependent on persuading two separate images to agree. It is still a selected company-affiliated demonstration: the post gives us no camera-calibration error, attempt count or test of continuity across an entire edited sequence.

Sources: George Kenwright / Google DeepMind / X [11]

What surfaced on X in our 24-hour sample

We searched X for Argon posts published from 13 October at 04:34 UTC to 14 October at 04:34 UTC. These are the highest-viewed relevant posts in the inspected sample, using visible counters captured around 04:35–04:45 UTC. Search ranking is selective, counters keep changing and views are not unique viewers. This is a documented sample, not an exhaustive ranking of X.

The useful editorial split is between access news, an imaginative result and a failure comparison. The launch post had the highest view count in our sample; the smaller filmmaking example answers a more specific question. Popularity helps identify what people noticed. It cannot tell us whether a model will preserve our own shot.

Sources: Google DeepMind / X [9] · Logan Kilpatrick / X [10] · fofr / Google DeepMind / X [12] · Semichenko / X [13]

The pelican is a small lesson in authorship

In fofr’s video, rotating Fourier epicycles resolve into a pelican riding a bicycle, then the view reveals the construction in depth. The creator, whose profile identifies Google DeepMind affiliation, credits Argon with the video and with prompting a time-synchronised Lyria 3.5 soundtrack. The audio therefore has its own model credit.

What makes it memorable is the reveal: the viewer first reads a drawing, then understands its mechanism. That is a directing decision. A code-based animation makes timing, camera position and construction parameters candidates for revision. Whether a model can execute the next note cleanly is the test that should follow the applause.

Sources: fofr / Google DeepMind / X [12]

The benchmark table has more than one winner

Google’s evaluation places Argon ahead on DeepSWE and long-context graph traversal. Astra leads this selection on FrontierSWE; Opus leads Terminal-Bench. These tests ask different questions. A high aggregate position cannot choose the best model for every step of a production.

Sources: Google DeepMind [3] · Google DeepMind [4]

Video understanding is not a cinematography score

The methodology matters particularly on LVBench. Gemini was evaluated at one frame per second; Astra used 800 frames, Fable 300 and Opus 600, reflecting API limits. Those are materially different inputs. Elsewhere, Google combines its own measurements with reported results and leaderboards; DeepSWE also uses model-specific agent setups. This is not one independent, identical test harness.

For a filmmaker, the next experiment should be narrower. Give competing systems the same footage and ask each to locate an eyeline mismatch, return the timecode and explain the evidence. Then check the actual frames. That would test a useful decision. It would still say little about pacing, performance or whether the scene deserves to exist.

Sources: Google DeepMind [4]

One million tokens is a budget, not a promise

The API exposes a shared 1,048,576-token window for input, thinking and output. Its default output budget is 65,536 tokens, including thinking. A long brief, references and intermediate reasoning all compete for space. “One million tokens” does not mean a guaranteed million-token answer on top of a million-token input.

To exceed 65,536 thinking-and-output tokens, the Interactions API requires background execution and a larger explicit budget. Google’s servers continue generation in hops of up to 64K tokens; a background interaction can run for up to 24 hours. Continuations reprocess context, billed at the cached rate when applicable. The production question is how much time and money it takes to reach an accepted result.

Sources: Google AI for Developers [5] · Google AI for Developers [6]

Then somebody pulled a gummy bear’s arm

Semichenko’s side-by-side recording compares interactive gummy bears made with Opus 5.5 and Argon. The published excerpt shows a conspicuous difference: the upper example has rounded, translucent figures, while the lower one develops a straight extended limb and overlapping figures. The author prefers Opus. This is a useful counterexample to the idea that a launch automatically settles visual coding.

The post reports $7.20 for Opus and $22 for Argon. We have not audited those runs or their settings. They cannot establish a general cost ranking. They do remind us to count retries and tokens alongside advertised prices. One attractive render and one disappointing simulation are both selected examples; neither is a reliable success-rate estimate.

Sources: Semichenko / X [13]

Where to try it, and what the API costs

Google lists Argon for Pro and Ultra users in the Gemini app, and Ultra users in Antigravity. AI Studio’s subscription-based Argon access requires Ultra; API billing is separate. Access therefore depends on the product surface, plan and applicable availability. The public rollout should not be read as free, unlimited access for every account.

Sources: Google [7] · Google AI for Developers [8] · Google AI for Developers [5]

The verdict: a strong candidate for the work between shots

Argon earns a place in a creative technical trial: scene construction, animation scripts, reference analysis and tools that keep a project editable. The launch evidence gives us reasons to test those workflows. It gives us much less reason to replace an established finishing process on the strength of a leaderboard.

Our proposed test is simple: build one room from two reference views, change the window position, then move one camera while keeping the room intact. Compare Argon, Astra and Opus using the same references, tools and acceptance criteria. Record total elapsed time, spend, manual corrections and whether the saved scene can still be edited. This is a test plan, not an experiment we have already run.

A numerical rating would imply more hands-on evidence than we have. The editorial judgement is firmer without one: Argon’s most interesting promise is that the visual result can come with a mechanism you can change. The director’s next note will tell us how much that promise is worth.

Sources: Google [1] · George Kenwright / Google DeepMind / X [11] · Semichenko / X [13]