Gemini 4 Carbon Builds Full Native Mac Apps From One Prompt
Gemini 4 Carbon one-shots signed, native Swift macOS apps from simple prompts, raising real questions for small software businesses.
What is Gemini 4 Carbon?
Gemini 4 Carbon is Google DeepMind’s newest frontier model, released as a companion to the earlier Gemini 4 model, Argon. Google describes the two as the same size, but positions Carbon as the smarter, slower sibling: Argon is built for speed, Carbon for depth. The headline capability, demonstrated by early testers, is that Carbon can take a short natural-language prompt and produce a complete, compiled, signed native macOS application in Swift, including features the user never explicitly asked for.
TL;DR
- Native app generation is now a real capability, not a party trick: Carbon produced a fully working, signed Swift macOS app from a single prompt, complete with a working Xcode project and a compiled binary ready to run.
- Unprompted feature design showed up in testing, including a menu bar controller for a Pomodoro app that the tester never asked for.
- App mashups worked too: given nothing but “research Bear and Obsidian and build me an app with the best of both,” Carbon produced a native note-taking app blending both products’ core ideas without any reverse engineering of the originals.
- Long-running, hands-off workflows are the intended use case. Carbon is built for tasks you kick off and walk away from, not quick back-and-forth chat, with reported median task times well above typical benchmark tasks.
- Pricing undercuts its closest rival: Carbon runs at a fraction of the per-task cost of Anthropic’s top model on independent benchmarking, even though it trails slightly in raw intelligence scores.
- Small software businesses face a new kind of competitive pressure as internal teams and individual users gain the ability to clone or approximate commercial SaaS tools for personal or in-house use.
How does Carbon generate native Mac apps?
In the demonstrations covered here, Carbon was given plain-English prompts inside an agentic coding environment (Anti-Gravity) and asked to build “native Mac apps.” It wrote the project in Swift using the standard macOS development stack, the same frameworks a human iOS/macOS developer would reach for, rather than wrapping a web app in a native shell. It then handled the build and signing process itself, producing both a full Xcode project and a compiled, runnable application.
For a Pomodoro timer app, the only real prompt was to build “the ultimate Pomodoro native Mac app” with useful features that “look fantastic,” plus a target folder. The result included standard timer functionality, but also audio cues for focus sessions, per-project tracking of completed Pomodoro sessions, and a menu bar controller so the app collapses into the Mac’s menu bar when minimized. None of those extras were explicitly requested. That detail matters: it suggests the model is drawing on a working internalized model of what a “good” macOS utility app looks like, not just executing a literal feature list.
Can it clone existing apps without reverse engineering them?
The second test pushed further: build an app combining the best of Bear (a polished note-taking app) and Obsidian (a linked-notes, markdown-vault tool popular for “second brain” style projects). The prompt was intentionally minimal: research both apps and build something combining their best features in a new folder.
Carbon did not access either app’s code or files. It worked from its training knowledge plus whatever it could look up, and produced a distinct native Swift app with a menu bar component and features recognizably inspired by both products. This is not decompilation or reverse engineering, it is closer to a capable developer studying two popular apps’ marketing materials, reviews, and screenshots, then building their own interpretation. The practical result is the same for a small business: a usable, native, free substitute for paid productivity software, generated in the time it takes to run a few meetings.
Why does self-checking matter for 3D and UI work?
Beyond native apps, Carbon was tested on a detailed, interactive 3D rendering of Marina Bay Sands in Singapore, a deliberately hard target chosen because it’s architecturally complex and likely underrepresented in training data compared to landmarks like the Eiffel Tower. The scene was mostly one-shotted, with only a handful of follow-up prompts asking for shader improvements and added environmental detail like clouds and people.
What stood out in testing was evidence of iterative self-checking during generation: the model appeared to revisit and refine details progressively rather than output a single flat pass. Google has talked about “self-verified UIs” as a design goal for Carbon, and the demo results are consistent with that: surrounding buildings near Marina Bay Sands, including distinctive shapes like the Apple Store’s sphere and the Louis Vuitton store’s pyramid, were rendered with enough fidelity to suggest the model was checking its output against real-world reference rather than guessing once and stopping.
Is Gemini 4 Carbon actually the smartest model available?
Not anymore, and that’s a useful data point for anyone choosing a model for production work rather than leaderboard bragging rights. When Carbon launched, it briefly took the top spot on the Artificial Analysis Intelligence Index. Within days, Anthropic’s newer model reclaimed the lead, scoring several points higher.
But intelligence score isn’t the only variable that matters. On the same independent index, the higher-scoring Anthropic model costs roughly three times more per task than Carbon. For most day-to-day engineering and knowledge-work use cases, that cost gap is arguably more decision-relevant than a five-point intelligence difference, especially for teams running many tasks at scale. The exception is frontier scientific or mathematical discovery work, where squeezing out maximum capability may justify the extra spend regardless of cost per task.
Carbon also runs notably slower and more verbosely than typical models on these benchmarks, generating far more tokens per task and taking roughly double the median time to complete. That reinforces its intended role: a model you assign a task to and leave running, not one you chat with interactively.
What does this mean for software businesses?
The practical implication of one-shot native app generation is uncomfortable for a certain category of software business: small, single-purpose SaaS products with high per-seat pricing and relatively simple feature sets. If a model can study an app’s public description, screenshots, and reviews, then generate a free, native, locally-run equivalent for internal use, the traditional moat of “we built it well and understand the problem” gets thinner.
That doesn’t mean all software businesses are at risk. Products with genuine network effects, ongoing data syncing, collaborative features, or deep backend complexity are much harder to replicate from a prompt. A likely defensive pattern is software companies pushing more of their actual logic and value into backend services in the cloud, keeping the client as a deliberately “dumb” interface, so that cloning the visible app doesn’t reproduce the thing that actually makes it valuable. Expect more businesses to lean into subscription-based cloud processing, proprietary datasets, or integrations that can’t be casually reconstructed by an agentic coding model, even a very capable one.
Frequently Asked Questions
What is the difference between Gemini 4 Carbon and Gemini 4 Argon?
Google describes both as the same underlying model size, but Argon is optimized for speed while Carbon is built for deeper reasoning and longer, more thorough task execution, at the cost of being slower to respond.
Can Carbon really build a complete, signed Mac app from one prompt?
In testing, yes: a single prompt describing a Pomodoro app produced a complete Swift-based native macOS application, including a working Xcode project, a signed and compiled binary, and unprompted extras like a menu bar controller.
Does Carbon copy existing apps’ code when asked to mimic them?
No. In the Bear/Obsidian mashup demo, Carbon worked from its general knowledge and research capability rather than accessing or decompiling either app’s actual code, producing an original implementation inspired by both.
Nothing gets lost.
Is Gemini 4 Carbon cheaper than competing frontier models?
On independent per-task cost benchmarking, Carbon runs at roughly a third of the cost of Anthropic’s top-scoring model, despite scoring a few points lower on raw intelligence measures.
Who is Carbon actually built for?
Google positions it around four use cases: long-running agentic engineering tasks, high-precision knowledge work like finance and legal analysis, frontier scientific discovery, and full-stack front-end or 3D web development, the last of which is where the native app and 3D rendering demos fall.