LeanBuddy budgets

LeanBuddy — the assistant in the LeanSignal app — is metered the same way the rest of the platform is priced: a transparent unit, published math, and a hard ceiling. This page is the entire cost model. There is nothing else, and nothing bills past the cap.

Credits

1 credit = 1 US cent of model spend. The meter in the LeanBuddy panel header shows credits used today against your daily allowance, each answer shows what it cost, and hovering the meter reveals the raw token counts and the model behind the numbers. Sub-cent usage displays as <1 — most simple questions cost less than a credit.

Cost-weighted tokens

Underneath credits, usage is counted in cost-weighted tokens: raw token counts scaled by the model provider’s published price ratios, so the same allowance represents the same spend no matter how an answer used the model.

Token class	Weight	Why
Input	1×	The baseline: your question plus the conversation so far
Cache write	1.25×	Storing a reusable prompt prefix costs slightly more once
Cache read	0.1×	Re-reading that prefix on later requests is 10× cheaper than resending it
Output	5×	Output tokens are priced at five times input on every model tier

These weights are price ratios, not tuning knobs — they mirror what the model actually costs, and they are stable across the model family. The practical consequences: LeanBuddy caches its fixed prompt aggressively (so follow-up questions in a conversation are cheap), and it answers tersely (because every output token is the most expensive kind).

Default daily allowances

Scope	Allowance per day	≈ on the default model
Per user	200,000 weighted tokens	≈ 20 credits ($0.20)
Per tenant (all users combined)	2,000,000 weighted tokens	≈ 200 credits ($2.00)

Both reset at midnight UTC. Allowances are denominated in weighted tokens, not credits, so they are model-independent: on a pricier model the same allowance simply represents proportionally more spend.

Hard caps — never overage billing

The caps are literal. When an allowance runs out, LeanBuddy stops answering until the midnight-UTC reset — it never keeps going and bills you later. There is no overage rate, no shared pool quietly draining, no surprise line item. The cap is enforced during a turn, not just at its start: a long investigation that reaches the ceiling wraps up its answer instead of running on.

This is lean by design, and it is the same principle the platform applies to telemetry: cost is bounded by an explicit declaration, not by whatever happened to be consumed. An AI assistant should never be the reason your observability bill surprised you.

Near the cap

LeanBuddy degrades before it refuses. When your remaining daily allowance runs low, the composer warns you, and answers switch to a terse mode — shorter responses, fewer telemetry queries — to stretch what remains. Only at the cap itself do turns stop until the reset.

The model

LeanBuddy defaults to a fast, cost-efficient model — the right tool for grounded questions over your telemetry, and the reason the default allowance goes a long way. Tenants that want deeper reasoning can be upgraded to a more capable model; the budget math above applies unchanged.

Why the allowance goes further than it sounds

LeanBuddy’s answers come from a store that is small by design: central storage holds only the demanded subset of your telemetry, so the assistant queries less, reads less, and spends less per answer than an assistant searching a collect-everything store. The condensed overview tools it prefers (error overviews, pattern mining, metric summaries) answer most questions in one or two queries. Specific questions with narrow time ranges cost the least — the per-answer receipt shows you exactly how each question spent.

The MCP server is metered elsewhere — by you

IDE and agent access over the MCP server (tokens are managed on the Preferences page) does not draw on these allowances: your own agent brings its own model, so the model spend happens — and is capped — in your own tooling. Heavy exploratory sessions belong there; the in-app assistant stays fast and bounded for everyone.

Related: Limits — the platform’s other numbers, same philosophy.