Google launches Gemini 4 Carbon with /swarm: swarm agents for long-horizon programming
Google released Gemini 4 Carbon on Tuesday (November 10) to paying developers via the Gemini API and to AI Ultra subscribers. The headline feature is /swarm, a preview capability in Google Antigravity that turns a single prompt into a coordinated team of AI agents, visible in real time.
What Gemini 4 Carbon is
Carbon sits above Gemini 4 Argon — the "standard" model for software engineering and knowledge work — and is designed for tasks that "test the limits of a model's intelligence," per the official announcement. Google positions it for long-horizon workloads: multi-package migrations, cross-subsystem debugging, complex root-cause investigation, and architectural redesigns.
Within Google's own 14,000-test codebase, Carbon achieved a 94% resolution rate on real bugs with zero regressions. On external benchmarks, it leads SEC-bench Pro at 88.2% (vulnerability discovery), ExploitGym at 54.8% (exploitation), and CWE-bench at 71.0% (patching) — results that substantially exceed Argon and place Carbon among the strongest models ever measured for cybersecurity.
The /swarm: programming in swarms
The most original feature is /swarm, launched in preview on Google Antigravity. Instead of a single agent coordinating isolated subagents — the pattern most platforms follow today — /swarm lets you orchestrate a multi-agent system where each agent can be "pinned" to a different model. The coordinator (Carbon) plans, workers build and verify in parallel, with a side-panel UI showing the Board, Roster, and Graph replay.
At Google, per the announcement, engineers used swarms for multi-package migrations, cross-subsystem root-cause investigations, and complex architectural redesigns. One example: with seven public NYC City Council planning PDFs (206 engineering drawing views extracted) and a single prompt, a multi-agent Carbon team used /swarm to reconstruct a 906K-triangle interactive Three.js mockup of Google's Platform 37 office in ten overnight loops.
Pricing is set at $2 per million input tokens and $10 per million output tokens through November 30 — matching Argon's introductory rate. After that, it doubles to $4 and $20. Google is offering six months of free access to 10,000 scientists (thousands already onboarded), including both Argon and Carbon.
Cybersecurity: the double-edged sword
Carbon was originally released in October only to defenders through the Fairwind Program. Since then, Wiz used the model in its Scan for Good program to identify zero-day vulnerabilities in two major SaaS providers in under two hours. In Google's internal Project Fortify (autonomous scanning, triage, remediation), Carbon hit 100% recall on scanning and 95% precision on triage, with over 25% relative improvement over Argon's already strong recall.
The model also reduced the per-attempt breaching rate to 0.8% in simulated environments against adaptive red-teaming — compared to 3.1% for Argon. Google acknowledges these results come from specific evaluation environments and don't establish resistance to all real-world attacks. The cybersecurity problem isn't solved — it's transferred.
The strategy behind the launch
Google's strategy with Carbon is clear: position a model that "costs more" but solves problems others can't, while feeding an ecosystem of tools (Antigravity, /swarm, Project Fortify) that creates natural lock-in for engineering teams. The six-month free offer to 10,000 scientists is both a generous gesture and a massive data-collection mechanism — every predicted protein, every identified vulnerability, every protein search query generates useful gradients.
The /swarm itself is the most significant signal: Google is betting that the next frontier isn't "a bigger model" but systems of models that work together. If this proves productive at scale, the question is how many other companies will copy this pattern before it becomes the new baseline for agentic programming.
Sources: Google DeepMind Blog, Engadget, The New Stack
✓ Independent sources cross-checked and verified before publishing
