Google releases Gemini 4 Carbon: tops three agentic benchmarks at less than half the cost

Google has released Gemini 4 Carbon, and the main emphasis is not on being "even smarter" but on being "cheaper and better for agents." The model is positioned for programming, AI agents, and scientific tasks. For those building products on agentic workflows, the most interesting parts are the price and the new parallel mode in Antigravity.

What Google claims

According to the company, Carbon took first place in three agentic benchmarks: CursorBench 4.0, Perplexity WANDR, and Zapier AutomationBench. These tests measure not "knowledge of facts" but the ability to carry a multi-step task through to a result: write code, execute a chain of actions, automate a workflow.

The second part of the claim matters more than the first: the cost of completing tasks is more than twice lower than that of the closest competitor. For agents this is critical. A single agent run means dozens of calls to the model, and the price difference per task quickly turns into a difference in product margin.

Google separately mentioned science: in experiments, Carbon helped find on average 40 times more remote protein homologs. These are proteins with a common origin but weakly similar sequences, and they are hard to find. The announcement gives no details on the methodology, so for now this is more a signal of direction than a ready-made tool for a lab.

/swarm: several agents instead of one

Alongside the model, Google showed the /swarm command for Antigravity. The idea is simple: instead of one agent that writes code, checks it, and runs tests in turn, several agents are launched in parallel. One writes, another reviews, a third tests.

Imagine the task "add a new endpoint to a service." A regular agent does it sequentially: code, then review, then tests, then fixes. In swarm mode, review and tests start while the code is still being finished. Waiting time drops, but the number of simultaneous model calls grows, and so does the bill. Here a low price per token becomes a condition for the mode to work, not a pleasant bonus.

Price and what to check before switching

Until November 30, the API costs $2 per 1M input tokens and $10 for output tokens. After that date, prices double: $4 and $20. The model is available via the Gemini API, Google AI Studio, and Antigravity.

For a rough estimate, take an agent run of 200 thousand input and 20 thousand output tokens. Right now that is about $0.60 (0.4 + 0.2); after November 30, about $1.20. The figures are illustrative, but they show the main point: the claimed "more than twice" advantage should be calculated at the future rate, not the promotional one.

What is worth doing in practice:

• Run your own tasks, not someone else's benchmarks. Leading CursorBench or AutomationBench does not guarantee the model will cope with your codebase and your integrations.
• Count the cost of a successfully completed task, not of a token. A cheap model that needs three attempts may turn out to be more expensive.
• Test /swarm on an isolated task and measure how much parallelism speeds up the work and by how many times token consumption grows.

Limitations

All results in the announcement are claims by Google itself. From the retelling, it is unclear exactly whom the cost was compared with, what settings the tests were run on, and what the price of a task is under real load. There are also no independent verifications. It is worth first waiting for third-party measurements or running your own pilot within the promotional window while the price is half as high.