Baseten
Connect your Baseten workspace to track billed spend across Model APIs, dedicated deployments and training jobs in StackSpend.
Baseten
Direct API billing — API key required
The Baseten provider connects StackSpend to your workspace through the management API’s billing usage summary. Baseten reports cost and usage together, broken into three product lines: Model APIs (per model, with the input, output and cached token split), dedicated model serving (per deployment, by instance type and environment) and training (per job). StackSpend normalises all three into one daily timeline alongside the rest of your AI stack, so a deployment left warm or a runaway batch job is visible the day it happens — not when the invoice lands.
What StackSpend tracks
- Daily billed Baseten spend in USD across all three product lines
- Per-model cost and token usage for Model APIs — input, output and cached input tokens
- Dedicated deployments by instance type and environment, with compute minutes and inference request counts
- Training jobs by job, with runtime minutes
- Credits applied to your workspace, spread across the spend they offset
- Anomaly detection and month-end forecast against your baseline
Prerequisites
- A Baseten workspace with billing
- A team API key created in Organization settings with the “Manage and call all team models” permission — Baseten has no billing-only permission level
Setup
Create an API key
Sign in at app.baseten.co and open Organization settings → API keys. Create a team key named “stackspend” — a dedicated key can be rotated or revoked without touching anything else you run on Baseten, and team keys survive a user being deprovisioned where personal keys do not. Copy it; Baseten shows it once.
Under Permissions, choose “Manage and call all team models”. The billing endpoint is not readable by the narrower levels — a key created with “Export model metrics” is rejected with a 403 PERMISSION_DENIED. Leave Restricted environment access empty: billing is reported for the whole workspace, so restricting environments gains nothing.
Add to StackSpend
Go to Providers → Add provider → Baseten:
Add Baseten provider
A team API key from Organization settings at app.baseten.co
Verify sync
Cost data appears within minutes of saving. StackSpend backfills daily history on the first sync, then refreshes recent days on each subsequent sync — Baseten updates deployment usage hourly, so the most recent days keep settling for a short window.
Data summary
| Attribute | Value |
|---|---|
| Granularity | Daily |
| Billed cost | Per model, per deployment, per training job |
| Token usage | Per model (input / output / cached input) |
| Historical data | Back to 1 January 2026 |
| Setup time | 2 min |
403 PERMISSION_DENIED on the billing endpoint. StackSpend only ever calls that one read endpoint and never writes to your workspace, and the key is encrypted at rest — but create a dedicated key so you can revoke it without touching anything else.Frequently asked questions
What credentials does StackSpend need to connect Baseten?
A single Baseten team API key, created with the “Manage and call all team models” permission. Baseten has no billing-only permission level — we tested the narrower ones, and “Export model metrics” is rejected with a 403 on the billing endpoint. StackSpend calls only the billing usage summary endpoint and never writes to your workspace.
How often does Baseten cost data update?
Daily. Baseten reports billed cost and usage at daily granularity and updates deployment usage hourly, so StackSpend refreshes the most recent days on each sync while they settle.
Does StackSpend show Baseten cost per model?
Yes, and the figures are your real billed amounts. Baseten reports cost and the input, output and cached token split on the same rows, so unlike providers that report dollars and tokens separately, no list-price estimation is involved.
Does StackSpend track dedicated deployments and training as well as Model APIs?
Yes. All three Baseten product lines are tracked on one daily timeline: Model APIs per model, dedicated model serving per deployment (including chains) by instance type and environment, and training per job.
How far back does Baseten history go?
To 1 January 2026. Baseten does not serve billing data from before that date, so StackSpend clamps the first backfill to it rather than requesting a range Baseten will reject.
Is the connection read-only?
StackSpend only ever calls the billing usage summary endpoint and never writes to your workspace. The key itself is not read-only, though: Baseten has no billing-only permission level, so a key that can read billing can also deploy and invoke models. Use a dedicated key so it can be rotated or revoked on its own.