Summary
Google says Gemini 4 Carbon is now available across developer, enterprise and consumer products, with access varying by subscription or programme. A preview feature in Google Antigravity coordinates multiple AI agents on larger projects.
Google announced Gemini 4 Carbon on 10 November 2026, making the model available through several Google products and programmes. The company describes Carbon as its most capable model for complex, multi-step work, including coding, cybersecurity and scientific reasoning. It also previewed /swarm, a Google Antigravity feature that coordinates multiple agents on a project.
Availability and the Antigravity agent team
Developers and enterprises can access Carbon through the Gemini API in Google AI Studio and Gemini for business on paid tiers. Google lists an introductory price matching Gemini 4 Argon: $2 per million input tokens and $10 per million output tokens. Carbon is also rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app and AI Mode in Search. The company lists additional developer access through Android Studio.
In Antigravity, /swarm turns one prompt into a coordinated team of agents that users can follow in real time. Developers can assign Carbon to demanding roles such as architectural design, coordination and auditing, while pairing it with faster, lighter models. The agents work across parallel, isolated workspaces, with adversarial verifiers checking changes.
Google’s examples illustrate the intended workflow. In one, a team used seven public city-planning PDFs containing 206 extracted engineering-drawing views to build an interactive 3D reconstruction of an office building. The company says the result had 906,000 triangles and was produced over 10 overnight loops. This is a demonstration of the system’s ability to divide and coordinate a project, rather than a general measure of how reliably it will complete other projects.
What Google reports about Carbon’s capabilities
Google says Carbon ranked first on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench, and reports more than twice lower cost per task than the next-best external model on those evaluations. For its new Multimodal Knowledge Agent benchmark, developed with Artificial Analysis to test agents working with dense documents and visual material, Carbon ranked second; Google says its average cost per task was more than six times lower than that of the top-performing model.
The announcement also highlights cybersecurity results. In Google’s internal Project Fortify benchmark, Carbon achieved 100% recall in scanning and 95% precision in triage. Recall measures how many relevant issues are found; precision measures how many flagged issues are relevant. Google says Carbon’s safeguards blocked 99% of attack trajectories in simulated cyberattack testing. The company also notes that prompt-injection robustness remains an area for continued improvement, despite reporting a 0% attack success rate on the Gray Swan indirect prompt-injection benchmark.
For scientific work, Google reports that researchers used Carbon in a multi-agent system to search large sequence and structure databases for remote protein homologs—evolutionarily related sequences that can help inform structure prediction. Across seven low-homology protein families from the CASP17 blind test set, the system retrieved 40 times as many homologs on average. Google says broader benchmarking is still underway, so this result is an early evaluation of a specific search task rather than a general measure of protein-prediction accuracy.
Carbon is positioned for work that benefits from extended reasoning and verification. Google says it uses more tokens than Argon for that deeper, longer-running process, while Argon remains its recommended choice for most software-engineering and knowledge-work tasks.