How does Baseten billing work?

Baseten meters dedicated deployments by the minute for every running replica at the instance's per-minute price, as of October 19, 2026: an H100 is $0.10833 a minute (about $6.50 an hour), an H200 $0.125 ($7.50 an hour), a B200 $0.16633 (about $9.98 an hour) and an L4 $0.01414 (about $0.85 an hour). Image builds, cold starts and model loading, serving and idle warm replicas are billed; image pulls, failed boots, replicas scaled to zero and development deployments are not. Partial minutes round up. Model APIs are billed per token.

Can I set a hard spending limit on Baseten?

Only partly. A monthly budget sends email alerts at 75%, 90% and 100%, and the Enforce budget option rejects Model API requests once spend reaches the budget. Enforcement never stops dedicated deployments or training jobs — they keep running and accruing charges. The real ceiling for dedicated inference is the max_replica setting on each deployment.

How do I control Baseten costs?

Set min_replica to 0 on anything that can tolerate a cold start, cap max_replica on every deployment, tune scale_down_delay (default 900 seconds) so replicas are not kept warm too long or churned too often, keep image builds lean because every push is billed, iterate on development deployments (not billed), and pull per-deployment usage daily from the billing API or CLI.

Does Baseten offer spot or discounted GPUs?

Baseten's public pricing does not list spot instances. Discounts are commercial: Pro and Enterprise plans offer volume discounts, and Enterprise adds on-demand flex compute, self-hosted deployments in your own cloud and the option to use existing cloud commitments. Ask for these in negotiation rather than expecting a spot price list.

Quick Answer

Baseten Billing and Cost Control: An Evaluation (2026)

Published: October 19, 2026

The short answer

Baseten bills each running replica by the minute, including the minutes it spends cold-starting, and its budget feature does not stop dedicated deployments. Cost control therefore lives in your autoscaling settings — min_replica, max_replica and scale_down_delay — not in the billing dashboard. Done well, a spiky workload costs a fraction of an always-on replica; done badly, a forgotten warm H100 costs about $4,680 a month. Facts below are from Baseten’s documentation, read October 19, 2026.

What you pay per minute

Instance	Per minute	Per hour (×60)	GPU memory
L4 (`L4:4x16`)	$0.01414	~$0.85	24 GiB
A100 (`A100:12x144`)	$0.06667	~$4.00	80 GiB
H100	$0.10833	~$6.50	80 GiB
H200	$0.12500	$7.50	141 GiB
B200	$0.16633	~$9.98	180 GiB
8× H100 (`H100:8`)	$0.86664	~$52.00	640 GiB

The Basic plan is $0 a month, pay as you go, with starting credits for new workspaces; there is no perpetual free tier. If credits run out and no payment method is on file, Baseten deactivates your models.

What is and is not billed

Lifecycle phase	Billed?
Image build after `baseten model push`	Yes
Image pull onto the node	No
Cold start and model load	Yes
Serving requests	Yes
Idle warm replicas (`min_replica` ≥ 1)	Yes
Replica that crashes mid-request	Yes, up to termination (partial minutes round up)
Failed boot (replica never came up)	No
Scaled to zero	No
Development deployments (`--watch`)	No

Two consequences people miss: every push costs money because builds run as billable workloads, and scale-to-zero is not free at the edges because each wake-up pays for the model load.

The cost arithmetic

One H100 kept warm all month (30 days): $0.10833 × 60 × 24 × 30 ≈ $4,680.
The same model used 8 hours a day on 22 working days, scaled to zero otherwise: $0.10833 × 60 × 8 × 22 ≈ $1,144, plus a few minutes of cold start each morning and after every idle gap longer than scale_down_delay.
Default scale_down_delay of 900 seconds means each burst keeps a replica billed for 15 more minutes after the last request. For bursty traffic that is often the largest line item.

The controls that actually cap spend

max_replica on every deployment. Default 1. This is the only hard ceiling for dedicated inference; above it requests queue or are shed.
min_replica: 0 wherever a cold start is acceptable. Baseten recommends 1 to avoid scale-from-zero latency and 2 if you also need redundancy — both cost money every minute.
Tune scale-down. Defaults: 60-second autoscaling window, 900-second scale_down_delay, at most 50% of replicas removed per step. Too short and replicas churn, each cycle paying a cold start; too long and you pay for idle headroom.
Set concurrency_target from a benchmark. Default 1 request per replica. An engine that batches well (vLLM, SGLang) can take far more, which means fewer replicas for the same traffic.
Budget alerts plus enforcement. Alerts at 75%, 90% and 100%; enforcement rejects only Model API requests.
Watch usage daily. Usage updates hourly; baseten org billing usage and the usage_summary API give per-deployment, per-day breakdowns (up to 31 days per query, no data before January 1, 2026).

When Baseten is not the cheapest option

You call popular open models and run nothing custom: a per-token API (Baseten’s own Model APIs, Together, Fireworks) avoids idle minutes entirely.
Very spiky, low-volume custom models: Modal bills per second; its Starter plan is $0 plus compute with $30 of free credits a month, and Team ($250 a month plus compute, $100 credits) adds environment-level budgets.
You already have committed cloud spend: negotiate Baseten Enterprise to run in your own cloud, or compare Baseten vs SageMaker vs Vertex AI.

For the full company evaluation see Baseten evaluation: pricing, strengths, alternatives; for packaging and migration traps, migrating AI inference to Baseten; and if an uptime guarantee is part of the decision, managed inference providers by uptime SLA.

Last verified: October 19, 2026. Per-minute prices from Baseten’s instance type reference; the monthly figures are our arithmetic on those prices.