
Tech • AI • Robotics • Game
Google has launched Gemini 4 Carbon, a new frontier model positioned as a major upgrade over Gemini 4 Argon, with stronger coding, long-horizon task execution, bug fixing and especially writing.
Gemini 4 Carbon is designed for two main use cases: cyber security and extended multi-step assignments that take significant time to complete. Google positions it as a direct successor to Gemini 4 Argon, and Carbon is described as outperforming Argon across all published benchmarks despite arriving only about a month later.
In hands-on testing inside Google Anti-Gravity, a coding environment similar to AI-assisted developer tools, Carbon produced a visually striking bioluminescent voice morphing engine that reacted to microphone input in real time. It also generated a Star Fox-like space fighter game with smoother controls, better ship physics and cleaner visual design than an Argon version, which appeared blockier and less responsive.
A top-down capture-the-flag paintball game built by Carbon showed sharper edges, smoother character motion and less blur than the Argon equivalent. The gap was not enormous on a simple project, but Carbon’s output was described as more refined and more playable.
When asked to create a portrait of an old man using only code, Carbon produced a sharper, more detailed result than Argon. But when tasked with recreating a supplied hiking photo purely through code, Argon initially looked better because it effectively embedded the source image as numeric data and reconstructed it, while Carbon tried to redraw the scene from scratch using mathematical instructions. That made Carbon’s first result less photorealistic but showed stricter adherence to the request.
Carbon was used to build FreeBooks, a bookkeeping app intended as an alternative to accounting software costing about $60 a month. The prototype handled manual transaction imports such as CSV files and was described as functionally competitive with paid tools, though automatic bank syncing remained unresolved. A recurring weakness was interface clutter: without explicit design guidance, Carbon tended to pack too much information onto a page.
Carbon was also tested on existing software built by other models. In one Mac-style app switcher, previous systems failed to fix a bug where minimized apps displayed only icons rather than preview windows. Carbon reportedly solved the issue in one pass, restoring previews for minimized applications and outperforming several competing models on a real debugging task.
Carbon’s strongest edge may be writing. It handled humorous fiction effectively and was judged especially useful for knowledge work where clarity and tone matter. In one workflow, Carbon drafted a new script using stored writing-style files, then edited a recorded presentation by removing pauses and mistakes while also choosing supporting visuals and generating motion graphics automatically.
The automated edit was described as highly polished, with accurate cuts and appropriate visual inserts, but it took noticeably longer than rival tools. The trade-off was speed versus a near-finished result in a single run.
Gemini 4 Carbon is rolling out globally first to Google AI Ultra subscribers in the Gemini app, Google Anti-Gravity and AI Mode in Google Search. Developers on paid tiers can access it through the Gemini API, Google AI Studio and Android Studio, while business customers can obtain it through Gemini for Business. Introductory API pricing through the end of November is $2 per million input tokens and $10 per million output tokens, rising afterward to $4 and $20.
Gemini 4 Carbon appears to be a meaningful step beyond Argon, with the biggest gains in coding reliability, bug fixing, sustained workflows and high-quality writing. Its launch strengthens Google’s position in advanced AI tools aimed at both developers and broader knowledge-work teams.
Ask a question