Gemini 4 Argon turns 80,000 NASA files into a deck, and Google shows 8 more tasks

Google has published eight examples of Gemini 4 Argon in action, led by a demo in which the model turned more than 80,000 raw NASA files into an executive-ready slide deck. Google describes Argon as its most intelligent model so far.

From a Mars data deck to a fur engine

The showcase is a set of Google’s own demonstrations, grouped by theme. The largest in scale involves NASA material: given a custom slide template and files covering weather logs plus engineering, health and science data from 72 flights on Mars, Argon running in Antigravity parsed the information, organized the findings and produced a deck with clear visual narratives, according to Google.

Other examples lean on code and 3D work. From a short prompt with no assets or libraries, Argon wrote 5,000 lines of math to render a sunlit 3D library and then made a video about the process. In Antigravity, it coded a custom fur engine with hair dynamics, and a self-navigating 3D periodic table whose animated atoms are timed to the lyrics of "The Elements" by Tom Lehrer.

Google also shows Argon coordinating other tools. In Antigravity it orchestrated subagents, Blender MCP and Nano Banana to turn photos of a real desk into a playable 3D track. In the Gemini app, it used Nano Banana and Veo to convert the concepts in an academic PDF into an illustrated animation, and it built a tool that traces musical scale lines across video, with motion triggering notes.

Who can use it today

Google says Gemini 4 Argon is globally available to developers and enterprises through the API in AI Studio and the Gemini platform. Google AI Ultra and Pro subscribers get it in Google Antigravity, in the Gemini app and in AI Mode in Search. The post gives no pricing, rate limits or per-region differences, so API costs and any usage caps remain unstated here.

Google’s separate post on five builders, dated October 23, adds outside projects such as a 3D seaplane simulation, a globe mapping NASA wildfire and USGS seismic data, and a multiplayer browser capture-the-flag game. It also does not publish benchmark scores or pricing.

Independent numbers are thin and mixed

The examples are curated and unmeasured: Google gives no success rates, run times, costs or failure cases, so how often Argon produces results like these cannot be judged from the post. Nor does it say how much human steering sat behind each demo, though the prompts are described as short in some cases.

One outside data point exists. In IBM’s ScarfBench Java migration leaderboard on Kaggle, which we covered earlier today, Gemini 4 Argon with Google’s Antigravity harness ranked first with a 75.2% smoke score across 204 tasks. Claude Opus 5.5 with Claude Code scored 70.5% and GPT-6.1 Sol with Codex 68.4%.

That lead comes with caveats. Claude Opus 5.5 with Claude Code posted the highest all-pass rate, 72.1%, and IBM says the entries are model and harness pairings, not pure model measurements. Because Argon was tested inside Antigravity, part of the gap may belong to the tooling, which is also the environment Google uses for most of its showcase demos.

For rivals, the competitive point is that Google is selling Argon as a system: a model plus Antigravity, Blender MCP, Nano Banana and Veo. Anthropic’s Claude Opus 5.5 recently arrived in Antigravity as well, billed pay-as-you-go, so users of that tool can already compare the two models in the same harness.

For readers, the practical change is the type of job being offered: long, multi-tool tasks that mix files, images and code, rather than single answers. Whether the NASA-style result holds on messy company data is something only trials on real files will show.