Google has announced its most powerful AI model yet, Gemini 4 Carbon. The model's standout new feature is Swarm: when a user issues a single command, the system automatically breaks the task into sub-tasks and runs multiple AI agents simultaneously. The feature is currently available only to subscribers of Google's top-tier paid plan, Ultra.

What happened?

Gemini 4 Carbon was introduced as the most capable model yet in Google's Gemini lineup. The Swarm feature takes a single prompt and splits it into sub-tasks such as research, code writing, and verification, with a separate agent instance handling each one. The agents then review and merge each other's output to produce a single final result. For now, the company is rolling out this capability not to the general public but to the Ultra subscription tier.

Why it matters

The idea of autonomous agents isn't new. In 2023, open-source experiments like AutoGPT and BabyAGI first demonstrated this concept, with a language model planning its own sub-tasks independently. But those experiments were fragile, often looping endlessly and producing nonsensical steps. In 2024, OpenAI also announced an experimental multi-agent framework for developers under the same name, Swarm. Google adopting that name for a consumer product today suggests the term has become an industry-standard concept. Anthropic, meanwhile, has pursued similar logic on the developer side through Claude's subagent architecture and its computer-use feature.

For Google, this move can also be read as defensive. The company has struggled to match ChatGPT's early lead in the chatbot race. While it has grown its user base through search engine integration and the Gemini app, in the 'agent' narrative it's been OpenAI's Operator and Anthropic's developer-focused tools that have stood out. Swarm looks like Google's attempt to appear as a pioneer in this space as well. But the real question is the reliability of the output agents produce: having multiple agents work in parallel doesn't automatically lower the error rate — on the contrary, it risks compounding each agent's own margin of error. The general experience across the industry has been that as the number of agents increases, so does coordination error.

Who does this affect, and how?

Swarm being locked behind the Ultra tier confirms a broader pattern seen across AI: the most capable features land on the highest-paying plan first, then trickle down to lower tiers. Under this model, free or mid-tier users may see no difference for months. For enterprise users, the real issue is cost; each agent instance consumes its own tokens, meaning a single Swarm command can be far more expensive than a single chat exchange.

CompanyAgent featureAccess
GoogleSwarm (Gemini 4 Carbon)Ultra subscription
OpenAIOperator / agent modesPro subscription
AnthropicClaude subagent architectureDeveloper/API, top-tier plan

The common thread in this table is clear: none of these companies is currently opening up autonomous agent capability to the free tier. That's a fact that cuts against the 'democratization' narrative often used to describe AI. All three major players are gating their most advanced capability behind a paywall — showing that the promised equality of access doesn't hold, at least not for this generation.

What's next?

There are three things worth watching. First, Swarm's error rate on real-world tasks — internal demos and independent user experience often diverge significantly. Second, when the feature will trickle down to free or standard tiers, which will signal how profitable Google considers it. Third, how competitors respond: whether OpenAI and Anthropic open up similar multi-agent capabilities to a wider audience will shape the pricing race ahead.

Swarm gives concrete form to AI companies' claim that we're moving from the era of the 'single chatbot' to the era of the 'autonomous team.' But whether that transition is reliable and accessible has yet to be proven; for now, Swarm sits at the intersection of an appealing promise and a limited audience.