What happened
Google announced Gemini 4 Carbon on November 10, 2026, describing it as its strongest model for frontier cybersecurity capabilities and long-horizon agentic work. The company distinguishes it from Gemini 4 Argon: Argon remains its suggested choice for most software engineering and knowledge work, while Carbon is aimed at complex, multi-step coding and debugging, open-ended scientific research, and tasks that benefit from deeper reasoning and extended verification. Google says Carbon uses more tokens than Argon for this more thorough process.
The model is available through the API in Google AI Studio and Gemini for business for paying developers and enterprises. 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. Google says it released Carbon to a group of cybersecurity defenders through its Fairwind Program two weeks before the broader availability announcement.
Google is previewing /swarm in Antigravity. The feature turns a single prompt into a coordinated group of agents that users can track in real time. Google says developers can assign Carbon to demanding roles such as architectural design, coordination, and auditing, while pairing it with faster, lighter models in parallel isolated workspaces and adversarial verifiers that test changes. As an example, Google says a multi-agent Carbon team used seven public city council planning PDFs containing 206 extracted engineering-drawing views to build a 906,000-triangle interactive Three.js reconstruction of its Platform 37 office over ten overnight loops.
Google reports that Carbon ranked first on CursorBench 4.0, Perplexity’s WANDR, and Zapier’s AutomationBench, at more than twice lower cost per task than the next-best external model. In a new Multimodal Knowledge Agent benchmark developed with Artificial Analysis, Google says Carbon ranked second and had more than six times lower average task cost than the top-performing model. For cybersecurity, Google reports results from its internal Project Fortify benchmark: 100% recall in scanning and 95% precision in triage. The announcement also describes early evaluations and examples from external testers, while attributing those results to the respective organizations or to Google.
What the archive adds
New Runtime's earlier coverage on November 3 described Carbon's cybersecurity and long-horizon coding focus, with access initially limited to cyber defenders and trusted partners. This is a new access event for the same model, not a new model announcement: Google now describes availability across developer, enterprise, scientific, and consumer surfaces.
Trend signal
This release supports agent economics moving to completed work: Google emphasizes benchmark cost per task, rather than token price alone, and presents Carbon as a model for tasks requiring extended verification. The comparison remains a vendor-reported benchmark claim, so it is a signal about how suppliers frame economics, not proof of lower cost in a team's own workflow.
Why now
Google says Carbon had been available to selected cybersecurity defenders through Fairwind for two weeks before this broader launch. The announcement marks a shift from that restricted evaluation to paid API access and rollout in consumer-facing Google products.
Why it matters
The distinction between Argon and Carbon gives teams a useful model-selection question: does a task require deeper reasoning and repeated checks, or will a faster, lighter option suffice? A lower token price or a favorable cost-per-task benchmark does not automatically mean a lower bill; Carbon's longer reasoning can consume more tokens. Teams should compare total cost and verified completion on representative work, not infer economics from headline rankings.
The /swarm preview also makes orchestration part of the product proposition. Parallel agents, isolated workspaces, and adversarial verification may help on large tasks, but add coordination and review overhead. The value will depend on whether the system produces reliable artifacts that people can inspect, and whether its verification loop catches consequential errors rather than merely generating more activity.
Limits and uncertainty
The benchmark rankings and cybersecurity results in the announcement are claims reported by Google; the announcement does not establish that they generalize to a team's own workloads. The Platform 37 reconstruction is a company example, not a controlled comparison of agent workflows. The announcement describes /swarm as a preview, so its operational reliability, limits, and performance at production scale remain to be established. Carbon's longer reasoning also means actual per-task API spend can vary with the task and token use, even at the listed rates.
Skill to build now
Build a small evaluation set from real work that separates tasks needing deep reasoning from routine coding or document work. Run both Argon and Carbon on the same tasks, capture input and output tokens, and require a human-checkable completion criterion. For any /swarm trial, record coordination time and the defects caught by its verification steps alongside final quality and cost.
Reader takeaway
Gemini 4 Carbon is now an option for paid API users and eligible Google product subscribers, with /swarm previewing coordinated agents in Antigravity. Treat Google's benchmark and cost comparisons as hypotheses to test: choose Carbon for work that benefits from deeper reasoning, then measure verified outcomes and total spend against Argon.