The cost model
The application is free. The cost is the models: a harness subscription, per-token API usage, or nothing at all if you use community shared compute. Hosting adds a cost if you self-host.
What changes with Buzz is not the price per token but the rate of consumption. A team of agents working in parallel across several channels consumes far faster than one person in one chat window, and nothing in the interface shows a running total.
The framing that works is token budgeting, not token maxing. Reported experience: today's $200 a month can become $2,000, and subsidised subscription tiers are widely expected to get more expensive.
Why it matters to you
For a team this is a budget question, not a technical one. An agent team that quietly becomes the largest line item on a project changes the economics of the project, and it does so without any single decision looking wrong.
The levers are all decided at configuration time, which is why understanding this before building a team is worth more than monitoring after.
How to apply it
Pin models per agent. This is by far the largest lever: a strong model for reasoning, a cheap one for summarising, formatting, routing and checking.
Cap parallelism on agents that do not need many concurrent sessions — though note the setting's effect is not fully verified; check your build.
Route high-volume low-stakes work to shared compute where the material allows it. See Buzz Mesh and where prompts go.
Stop loops quickly; a runaway thread bills for as long as it runs. See Stop, steer or recover a runaway agent.
Set a review point and a rough ceiling you would want to know about crossing, since nothing will tell you.
Treat any in-app cost figure as labelled rather than exact. The metering rules require "price unknown" whenever a request is routed through a gateway, a custom base URL or an unresolved alias, and forbid merging estimated with reported costs into one unlabelled total. A number without provenance is not a number you can bill or budget against.
Reserve multi-agent threads for work where challenge changes the answer. Two agents agreeing is the expensive way to confirm your first idea.
If you ignore this
Left unmanaged, cost is the most likely reason a promising agent-team setup gets shut down — not because it failed, but because nobody could explain the invoice.
Managed with model pinning and a monthly review, the same team is affordable and its cost is attributable to projects.
Examples
Staffing a project. You do not put your most expensive person on formatting, and you notice quickly if someone does.
A five-agent team where only the analyst runs a premium model. Most of the token volume goes through cheap or free compute, and the quality where it matters is unchanged.
Setting every agent to the strongest available model "to be safe" and raising parallelism to twenty for speed. Both defaults are individually reasonable and jointly expensive.
Verified against Buzz v0.5.20 · Updated 2 Sep 2026