Google has made Gemini 4 Carbon available to paying developers and businesses through its API, and to its top-tier Ultra subscribers in the Gemini app, Search and its coding tool Antigravity. Free and cheaper plans do not get it. Alongside it, Google previewed /swarm, which starts a team of agents from one request and lets you watch them work. The question for anyone planning to use a team like that is what stops it from producing confident mistakes in bulk.
A team makes more output than anyone can read
Antigravity could already hand parts of a job to helper agents running in parallel. Swarm adds the whole team from one prompt and a live view of it. Google suggests giving Carbon the roles that need the deepest thinking, such as design, coordination and auditing. Lighter, cheaper models do the bulk of the work, and a separate set of agents tries to break each change.
That is how a firm works: a senior person reviews what juniors produce. The difference is volume. A team of agents can produce more in an hour than one person can read in a day, so the review has to happen automatically or it does not happen.
The Signal proof is the strongest kind of check
The Signal work shows what an automatic review looks like at its best. A sudoku works the same way: finding the answer is slow, but checking a finished grid takes a minute, and you do not care how the solver got there. In the Signal proof, the agents were the solvers and the Lean proof checker was the person checking the grid. To make sure the checker really caught errors, the researchers put two real, previously fixed bugs back into a copy of Signal's code. In both cases the system produced a proof that the buggy version was insecure.
People still did the part that needs judgment. Human experts decided what security properties had to be proven, and the agents then did the heavy work. That split, people choose the target and a machine checks the result, is the one to copy.
What this looks like in an ordinary office
Take a finance team closing the month with 2,000 supplier invoices to match against purchase orders. Cheap agents match most of them. A program then checks rules that cannot be argued with: every matched invoice points to an order that exists, no invoice is paid twice, and the matched total equals the ledger. Perhaps 200 invoices fail and need real thought. Sending those to Carbon, at about 10,000 tokens in and 2,000 out each, costs roughly $8 at the current price of $2 per million input tokens and $10 per million output tokens. A person spending ten minutes on each would need about 33 hours, or four working days. The person still decides which exceptions to pay.
The price is introductory. It doubles at the end of November, so the same job costs about $16 afterwards. Carbon is also built to use more tokens than Argon, Google's cheaper model from a month earlier, so the lesson from our Argon pricing piece applies again: a task costs what it uses, and the price per token only tells you the rate.
Checks made of models are weaker than checks made of rules
Swarm's verifying agents are themselves models. A model that reviews another model's work can share its blind spots, so it lowers the error rate but cannot certify anything. A rule such as "totals match" or a test suite that passes cannot be talked into agreeing. My expectation is that the first jobs given to agent teams in ordinary companies will be the ones that already have a rule-based check: software with tests, reconciliations, and reports that must follow a fixed format.
The practical step is to ask what would tell you a result was wrong before you hand a job to a team of agents. If a program can ask that question for you, the job is ready. If only you can, the team will mostly hand you more reading.