Costs

Runs are billed by the amount of compute they use, measured at run time. Before you submit, you can ask the API for a cost estimate — and while you're building an integration, you can run end-to-end for free in test mode.

Estimate a cost before running

Every job type has an estimate-cost endpoint that returns an estimate without creating a run or using any GPU:

• POST /compute/v1/predictions/structure-and-binding/estimate-cost
• POST /compute/v1/predictions/adme/estimate-cost
• POST /compute/v1/protein/design/estimate-cost
• POST /compute/v1/protein/library-screen/estimate-cost
• POST /compute/v1/small-molecule/design/estimate-cost
• POST /compute/v1/small-molecule/library-screen/estimate-cost

The CLI exposes the same estimate commands:

Job type	CLI command
Structure and binding prediction	boltz-api predictions:structure-and-binding estimate-cost
ADME prediction	boltz-api predictions:adme estimate-cost
Protein design	boltz-api protein:design estimate-cost
Protein library screen	boltz-api protein:library-screen estimate-cost
Small-molecule design	boltz-api small-molecule:design estimate-cost
Small-molecule library screen	boltz-api small-molecule:library-screen estimate-cost

Pass the same input you'd use to start the run:

est = client.small_molecule.design.estimate_cost(
    target={"entities": [{"type": "protein", "value": "...", "chain_ids": ["A"]}]},
    num_molecules=100,
)
print(est.estimated_cost_usd)        # e.g. "0.5000"
print(est.breakdown.num_units)       # 100

The response encodes money as decimal strings in USD (to preserve precision):

{
  "estimated_cost_usd": "0.5000",
  "breakdown": {
    "application": "small_molecule_design",
    "num_units": 100,
    "cost_per_unit_usd": "0.0050"
  },
  "disclaimer": "This is an estimate only and may differ from your actual charges."
}

estimated_cost_usd is the authoritative total. num_units is what you requested — samples for structure-and-binding, molecules or proteins for design and screens. cost_per_unit_usd is a rounded per-unit figure and may bundle overhead, so it won't always multiply out exactly to the total.

The numbers above are an illustrative format, not a price quote. The estimate is approximate — final billing is based on the exact compute measured at run time, and for large library screens the estimate is extrapolated from a sample, so it's less precise for highly variable inputs.

Develop for free in test mode

A test-mode API key (mode: "test", prefix sk_bc_*_test_) runs the full request → poll → fetch-results lifecycle against the real endpoints but returns synthetic results with no GPU usage and no billing. Use it to build and test your integration for free, then switch to a live key for real runs. See Test environment and key prefixes.

Track usage and cap spend

Retrieve aggregated usage with admin.usage.list (GET /compute/v1/admin/usage, admin key required). You give a time range and window_size (HOUR or DAY) and optionally group by workspace or application:

for bucket in client.admin.usage.list(
    starting_at="2026-05-01T00:00:00Z",
    ending_at="2026-05-31T00:00:00Z",
    window_size="DAY",
    group_by="workspace_id",
):
    print(bucket.start_time, bucket.workspace_id, bucket.quantity)

Each bucket's quantity is the aggregated billed unit count for that window (not a dollar amount).

To cap spend, set a per-workspace spending limit (a lifetime cap, denominated in milli-USD) on the workspace's admin endpoint. Workspaces with no limit set bill against the organization. See Organizations & workspaces.