Models and costs

You bring your own model key and pay your provider directly. Debutly never sees your bill.

What a run costs

Every run prints an estimate before it starts:

Estimated cost: $0.03–$0.05 (3 agents × up to 30 steps on z-ai/glm-5.3-flash)

With the default models, a quick run of three agents usually costs a few cents. A business crew run on a real SaaS app (2 owners, 4 teammates) cost about $0.05. The report says what each run actually cost, and npx debutly costs shows your OpenRouter key's spend by day, week and month.

Keeping spending in check

• maxCost refuses to start a run whose estimate could exceed it. The GitHub Action sets it to $1.00 unless you say otherwise.
• A key limit. Give your OpenRouter key a monthly limit. Before each run Debutly reads what is left on it, refuses a run that can't be paid for, and warns when one might run short.
• Running out part-way. If the provider refuses for lack of credit during a run, every agent stops at its next step instead of each paying for a failed request, and the report says the run was cut short. On a low balance OpenRouter can refuse a request for the most it might cost; Debutly then asks again for a reply the balance covers.
• Per-agent caps: maxSteps, agentTokenCap and agentMinutes.

The models

With an OpenRouter key, Debutly uses:

• z-ai/glm-5.3-flash for the agents, with low reasoning effort. It found as many seeded bugs as Gemini 3.8 Flash for about a quarter of the cost, taking about 40% more steps, which is why maxSteps defaults to 30.
• deepseek/deepseek-v4-flash to cast the customers and judge pages.

Change them with DEBUTLY_NAV_MODEL and DEBUTLY_JUDGE_MODEL. Any model on OpenRouter works, and so does any OpenAI-compatible provider (DEBUTLY_PROVIDER=openai-compatible and DEBUTLY_BASE_URL) or Anthropic (DEBUTLY_PROVIDER=anthropic).

Models compared on the seeded-bug fixture (5 planted bugs, 3 agents), September 2026:

google/gemini-3.8-flash is the fastest and needs the fewest steps; DEBUTLY_NAV_ROUTING='{"order":["google-ai-studio/flex"]}' halves its price, but Google AI Studio isn't among the providers that keep no data, so that also needs DEBUTLY_ZERO_RETENTION=off (see Data and privacy). To compare models on your own app, see scripts/compare-models.ts in the repository.