Cost forecast

Cost before the first token

The builder forecasts what a run will cost while you design it — no API calls, nothing executed. Most visual builders only show cost after a run, on their key. Breakerbox shows a p50–p95 band per node and for the whole graph, right on the canvas.

An estimate, not a quote. It's a heuristic — clearly labeled as one — and gets calibrated by your own run receipts over time.

Reading the forecast

• The band. p50 is the typical run; p95 is the bad-but-plausible tail — the loop that spins. A single number hides that tail, which is the whole point of a budget tool, so the forecast is always a range.
• Per-node badges. Each model node shows run ≈ $p50–$p95. A node inside a loop is marked with a brass ↻ — its cost is multiplied by the expected iterations.
• Total vs budget. The Budget Tree shows the total band against your root budget. If the p95 tail could exceed the budget you get a warning — not a block. The hard export gate stays the sub_budget over-allocation check.

How it's computed

Expected cost = per-call cost × expected calls, summed across the graph:

• Per-call cost comes from the price table (~1000 input tokens + max output at the model's rate).
• Expected calls is a static walk of the graph — routers split flow across their branches, and nodes inside a cycle are multiplied by a loop-iteration assumption.
• The band comes from two honest sources of variance:
  • loop iterations — p50 few, p95 many;
  • output tokens — p50 a fraction of max_tokens, p95 the full cap (what the guard reserves).

What-if

• Loop slider. Drag "loops ≈ N×" and the whole band and every node badge recompute live — watch a runaway loop turn a $0.40 run into a $3+ tail.
• Model swap. On any model node, pick a different model under "try a different model (preview only)" to see the whole-graph forecast recompute — e.g. swapping a branch to Haiku might cut it −67%. The preview never changes your graph.

It never touches your code

The forecast is advisory. It's computed entirely in the browser and does not affect the generated Python — the code you export is exactly what the golden-fixture tests lock, forecast or no forecast.