Home / Vibe Game / AI tools / Gemini
Gemini 4 Carbon for game makers: swarm, pelican demo, 3D claims
Carbon is pitched at long agentic runs, and Google says its 3D and creative coding work carries into games. This walks through the model's place next to Argon, the /swarm command, a playable pelican campus demo and where to try Carbon.
What Google says Carbon is
Google released Gemini 4 Carbon on November 10, 2026, through the Gemini API, Google AI Studio and its consumer and enterprise products. Google's announcement calls Carbon its strongest model, with frontier performance in cybersecurity and in long-horizon agentic workflows across knowledge work, software engineering and scientific reasoning. Those are the company's own characterizations, and they are framed around general workloads, not around games.[1]
Argon, released shortly before Carbon, remains Google's recommendation for most software engineering and knowledge work. Google positions Carbon for tasks that test the limits of a model's intelligence, such as multi-step coding and debugging and open-ended research, and says it spends more tokens than Argon to explore, reason and run verification loops. The developer documentation draws the same line: both models share a 1M-token context window, built-in tools and per-token price, and the practical difference is how many thinking tokens each spends per turn, which raises token use and latency on Carbon.[1, 2]
3D game development and creative coding claims
Google's post says stronger visual design, spatial reasoning and attention to environmental detail carry over to 3D game development and creative coding. It also says Carbon runs build-and-verify loops for web apps, recovers from errors and checks both its code and the rendered UI. These are qualitative statements: the post gives no game-specific measurements for the 3D claim, with no build times, asset counts or performance figures attached to it.[1]
The figures Google does publish are general model benchmarks run by Google and its partners. According to Google, Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at more than 2X lower cost per task than the next external model it compares against, and second on the Multimodal Knowledge Agent benchmark that Google built with Artificial Analysis, at more than 6X lower average cost per task than the top model there. None of those benchmarks measures a game.[1]
The pelican campus game Google published
Google built a browser game in which you explore the Mountain View campus as a pelican riding a GBike, and published it as a live app. According to the post, the build used the /swarm feature in Google Antigravity to coordinate subagents on separate pieces: modelling 3D assets in Blender from Omni and Nano Banana references, building the browser game in Three.js with WebGL shaders, and pairing procedural Web Audio sound effects with Lyria music.[1]
The post gives no numbers for this demo, and it does not describe the pelican game as coming from a single prompt. Google uses that phrase for two other builds in the same post, an Eiffel Tower construction exhibition and an office reconstruction. The demo runs at https://aistudio.google.com/apps/bundled/pelican-campus-quest, and opening it requires signing in with a Google account.[1]
How /swarm works in Google Antigravity
The /swarm command turns the current conversation into a coordinator that splits a task and adds a named worker for each piece, for example one worker per file or per package. Each worker starts with a clean context and a public brief posted to a shared board, stating what it owns, what it must not touch and what done looks like. The coordinator usually adds a verifier for every two to four workers, and that verifier reads diffs, runs tests and reports problems to the worker that wrote the code. Workers do not message each other privately, idle members sleep on the board without using tokens, and the coordinator hands out tasks and judges what comes back while the workers run.[3]
The docs list /swarm on Google AI Pro and Google AI Ultra plans, in both the Antigravity App and the Antigravity CLI, and mark the feature as in preview. Because every worker is its own agent, a swarm uses quota much faster than a single conversation, according to the docs. A member that runs out of quota pauses instead of retrying and stays paused until a new post appears on the board, its model is switched, or you message it.[3]
Plans and where to try Carbon
Gemini 4 Carbon is not available on the free tier of the Gemini API, per the documentation. Google lists an introductory price that runs through November 30, 2026, with a higher price applying from December 1, 2026. The current amounts sit in the data table on the Gemini tool page.[1, 2]
Developers reach Carbon through the paid Gemini API tier, in Google AI Studio and Android Studio, using the model ID gemini-4-carbon. Google AI Ultra subscribers can use it in Google Antigravity and AI Studio, enterprises through Gemini for business, Google Antigravity or CodeMender, and everyday users through the Gemini app and AI Mode in Search. Google also says current and future participants in its scientist access program can use Carbon.[1, 2]
Rewritten from the official pages listed below, indexed 2026-11-14. Benchmark and performance figures are reported by the company and its partners. Prices and availability may change; the official pages take precedence.
Sources
- blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-carbon-now-availableindexed 2026-11-14
- ai.google.dev/gemini-api/docs/whats-new-gemini-4-carbonindexed 2026-11-14
- antigravity.google/docs/swarmindexed 2026-11-14
