Google releases Gemini 4 Carbon for heavy-duty agentic workflows

If Gemini 4 Argon was Google's dependable daily driver for routine coding, the newly released Carbon is the heavy machinery rolled out for the toughest structural problems.

Following a brief two-week preview limited to vetted cybersecurity defenders, Google officially launched Gemini 4 Carbon for paying API developers, enterprise clients, and Google AI Ultra subscribers. Google positions it as its most capable model to date, purpose-built for sustained, multi-step agentic workflows that test the upper limits of machine reasoning.

Deep reasoning and the /swarm preview

Google is very explicit about the division of labor. For writing day-to-day features or iterative script editing, Argon remains the faster, cheaper choice. Carbon is designed for deep-water tasks: multi-package code migrations, cross-subsystem root-cause debugging, formal mathematical verification, and open-ended scientific discovery.

To make that capability usable on complex projects, Google introduced a preview command in Antigravity called /swarm. A single prompt can now spin up a coordinated crew of specialized agents. A developer can designate Carbon as the chief architect and auditor to oversee the plan, while routing smaller implementation tasks to faster, lighter models in isolated workspaces, all double-checked by adversarial verifiers.

On the capability front, Carbon puts real muscle into cybersecurity and rigorous logic. Internal teams used it alongside Wiz to uncover zero-day vulnerabilities in under two hours across major SaaS providers, while researchers used it with AlphaProof Code to formally verify cryptographic code for Signal in Lean. Crucially, it hallucinates far less under pressure: on an internal 375-task agentic honesty test, Carbon passed 92% of the tasks, declining to fabricate missing data 98% of the time.

Competing on unit economics

Carbon's rollout lands squarely against Anthropic's new Claude Fable 5.2. On the Artificial Analysis leaderboard, Fable 5.2 edged out Carbon 65 to 60. But instead of chasing benchmark vanity points, Google is attacking competitor pricing.

On Zapier's AutomationBench, Carbon took the top spot while logging a cost of $2.12 per task—less than half of Fable 5.2's $4.64. Similarly, on the Multimodal Knowledge Agent (MKA) benchmark, Google claims Carbon runs at more than six times lower cost per task than the top performer.

Access, however, comes at a steep price. The standard $19.99 monthly Pro plan does not include Carbon. To use it in the Gemini app, Search AI Mode, or Antigravity, consumers must upgrade to Google AI Ultra, starting at $99.99 per month. For API developers, an introductory rate of $2 per million input tokens and $10 per million output tokens is valid until November 30, after which prices double to $4 and $20.

Why it matters

Google is making a smart strategic pivot: when frontier intelligence scores begin to bunch up, the winning pitch is production economics and multi-agent ergonomics.

Features like /swarm show that Google understands where advanced AI is actually heading. Solo frontier models are too expensive and slow to handle whole projects alone; splitting architecture and verification to Carbon while delegating grunt work to cheaper models is the right operational pattern.

Still, regular users should tread carefully. The $99.99 starting price makes it clear Carbon is strictly a prosumer and enterprise tool. And while Google emphasizes low per-token rates, Carbon's extended verification loops and multi-agent swarms mean real-world token burn can add up fast. It is well worth testing during the promotional window for heavy architectural or security workloads, but routine developer tasks are still best left to Argon.