Google lança Gemini 4 Carbon com swarm de agentes e preço agressivo por token

Google began rolling out Gemini 4 Carbon to the general public on November 10, 2026, after two weeks of restricted access to a group of "cybersecurity defenders." According to the company, it is the "strongest" model it has ever released, designed to compete directly with the cutting-edge products from Anthropic and OpenAI. The news was reported by The Verge.

The launch comes just over a month after Gemini Argon, another frontier model from Google that arrived in October 2026. For those following the big tech AI race, the move confirms something that had already been taking shape: after months in which the competition seemed limited to Anthropic and OpenAI, Google returns to the front line with two heavyweight launches in a row.

Carbon vs. Argon: Two Models, Two Audiences

Google was direct about when to use each model. Argon remains the recommended option for most day-to-day software engineering and knowledge work. Carbon is reserved for tasks that test the limits of the model's intelligence, such as complex multi-step coding and open-ended scientific research.

Built for tasks that test the limits of a model's intelligence.

Google, in official blog post

In practice, this means engineering teams don't need to switch models across the board: Google's own recommendation is to keep Argon as the default and bring in Carbon selectively, for specific higher-complexity tasks, such as large refactors, intricate data pipelines, or research problems without an obvious solution.

Swarm: One Prompt Becomes a Team of Agents

The functional highlight of the launch is the "swarm" feature, available in preview on Google's agentic development platform, Antigravity. The idea is to turn a single prompt into a coordinated team of agents working on the same task, instead of relying on a single agent executing step by step.

For those already running agents in production, the multi-orchestration model changes the cost math directly: more agents running in parallel tend to consume more output tokens overall, even if each individual call is cheaper. This is worth testing before migrating entire workloads to swarm, since any gain in speed or quality can be offset by the volume of tokens generated by agents coordinating with each other.

Aggressive Pricing, but with an Expiration Date

As it did with Argon, Google is coming in aggressive on price. Carbon debuts with an introductory rate of $2 per million input tokens and $10 per million output tokens.

Fable 5.2, from Anthropic, was launched on November 9, 2026, one day before Carbon's public rollout. In the benchmarks released by Google itself in the Carbon announcement, the model appears as one of the strongest on the market across most metrics, trailing only Fable 5.2. The usual caveat applies: benchmarks released by the manufacturer tend to be cherry-picked to showcase the product in the best possible light.

Carbon's introductory rate has a set expiration: prices double at the end of November 2026, rising to $4 per million input tokens and $20 per million output tokens. Even after the adjustment, the model remains significantly cheaper than competitors per token, according to Google. In short: for those finalizing AI infrastructure budgets for December, it's worth running the numbers now, while the launch pricing still applies, and reviewing them again after the increase.

A technical detail that per-token pricing doesn't capture on its own: different models can use different amounts of tokens to solve the same task. A model that's cheaper per token but generates longer responses or requires more intermediate steps in an agent can end up costing more overall than a rival that's more expensive per token but more economical in tokens generated. This is especially relevant with swarm, which by nature multiplies the number of calls.

Where Carbon Is Available

The model is already available via the Gemini API in Google AI Studio and in Gemini for Business, for all developers and companies on paid plans. Google AI Ultra subscribers also have access through the Gemini app, AI Mode in Google Search, and Antigravity.

Google has not said whether, or when, Carbon will reach subscribers on lower-tier plans or free users. For those currently maintaining production products with Argon or with models from other providers, this means that testing Carbon in a real environment, for now, requires being on a paid tier.

What Remains Open

Google did not detail a timeline for swarm to exit preview on Antigravity, nor when (or whether) the feature will become a stable functionality outside the testing environment. There is also, as of the original source's publication, no independent comparison of total cost (tokens consumed per task) between Carbon, Argon, GPT-6 Astra, and Fable 5.2, which leaves the decision of which model to use in each case dependent on each team's own testing before any production migration.