DEV Community

Rose
Rose

Posted on

What a crew of agents costs against one big agent — the arithmetic, without invented numbers

The first question anyone asks about running agents in production is not "which model" — it's "what will this cost me". So let me answer it the way I'd want it answered: with the mechanism, the numbers we actually publish, and no invented token spend.

The expensive default

The cheapest way to build an agent is one model, one long conversation, everything in the context window. The problem is arithmetic, not design taste: in a chat-style loop, the whole history is re-sent every turn. Turn 5 pays for turns 1–4 again. Turn 40 pays for 39 turns of context to make one small decision. Your instructions, your tool schemas, your file contents, the whole transcript — billed again, each turn, whether that turn needed any of it or not.

You feel this as a bill that scales with how long the conversation got, not with how much work was done. The work was one paragraph. The invoice is the transcript.

What a crew changes

Split the job across specialists and three things stop happening at once.

Each specialist carries a short context. The one writing the tests never sees the deployment log. The one reading logs doesn't hold the schema. The lead holds the plan and the summaries that come back, not every intermediate step. Less context per turn is less billed per turn — the same reason you don't paste the whole codebase into a chat to rename a function.

Each specialist can run a different model. This is the lever most people leave on the table. The specialist doing mechanical extraction doesn't need a frontier model; the one reasoning about an ambiguous requirement does. Per-agent model selection means you pay frontier prices for the part that needs judgement and cheap prices for the part that doesn't. One big agent with one big model cannot make that trade — every turn is priced at the top of the range.

Idle agents make zero calls. No heartbeat, no polling, no "still thinking" turn burning tokens while it waits for a human. An agent wakes when there's work, does it, and stops.

The honest caveat about caching

Prompt caching is the other big lever, and it is easy to overclaim. Caches live in minutes, not days. The claim we will stand behind is narrow and true: prompts are built so the cached prefix stays byte-identical during an ongoing conversation, which is when caching actually pays. A scheduled job that runs every six hours does not run at cached rates — nobody's does. If someone tells you their nightly schedule is cheap because of cache, they're describing a cache that expired hours ago.

The part we charge for

The plan price is flat and per-account, and your model provider bills you directly at cost — we sit in the middle for orchestration, not for tokens, and take no markup on them. The plans differ in what they let you run at once, which is the honest shape of the cost:

Free Solo $24 Entrepreneur $69
Agents 2 5 25
Chats running at once 1 3 10
Context per run 100k 400k full model window
Schedules 1 10 50
Shortest schedule gap 6h 15m 5m
Sandboxes 0 2 3

Read that table as a parallelism budget, because that is what it is. Concurrency is what you're actually buying: ten chats at once will outspend one chat, by design. Parallelism is the purchase, not a discount.

What I will not do

I won't quote you a dollars-per-month token figure. It depends on your model, your workload and how much your agents talk to each other, and any single number I gave you would be a guess dressed as a measurement. What is measurable from our side is per-specialist spend, live, in the console — you can see which agent is eating the budget rather than discovering it on the provider dashboard.

Measure it yourself in an afternoon

Pick one real job you already run. Run it once as a single agent in a long conversation, then once as a crew — a lead plus two specialists with short contexts and different models. Compare the provider's own usage page for the same output. That comparison takes an afternoon and it is worth more than any benchmark, mine included.

The mechanism is the argument: short contexts, per-agent models, no idle calls. If the arithmetic doesn't hold on your own workload, no pricing page will save it.


AgentParley is hosted crews of AI specialists — a lead splits the work, specialists report back, schedules wake them on time, and your model provider bills you at cost. If you're weighing it against running your own, the honest comparison is here; the rest is at agentparley.ai.

Top comments (0)