Baseten

What it is for

Baseten is a production inference cloud: call a hosted model through an OpenAI-shaped API, or deploy your own checkpoint on a dedicated GPU that scales to zero. Use it when the job is serve the model, not when you need an SSH Pod.

It is not fully closed. Truss packages the model. Hosted inference is paid. New accounts include trial credits. Dedicated H100 is $0.10833/min (~$6.50/hr).

What makes it unique: the cover is Baseten’s official Model APIs price table — GLM / DeepSeek / Kimi K3 $3.00 / $0.30 / $15.00 per 1M plus Try Model API.

Key features

• Model APIs — hosted open models on an OpenAI / Anthropic-compatible endpoint (Quickstart).
• Dedicated deployments — pin a custom or fine-tuned model to GPUs you size (Pricing).
• Truss — open-source package format for any framework (Truss).
• Scale to zero — you do not pay idle time; cold starts are billed while the replica comes up (Pricing FAQ).
• Training jobs — on-demand VMs / GPUs on the same price table.
• Self-host / VPC — Enterprise can run Baseten in your cloud (Pricing).

Who it's for

Best fit: ML teams who will call one inference URL and pay per token or per GPU-minute.

Poor fit: researchers who want a cheap SSH box for a weekend (Runpod).

Pricing

Verified 2026-10-22 (Asia/Shanghai) from baseten.co/pricing. Plan names (Basic / Pro / Enterprise) gate support and rate limits; dollars are usage. Re-check the model row.

Plan cards — Basic includes dedicated deployments, Model APIs, training, SOC 2 Type II + HIPAA, email / in-app chat. Pro adds priority GPUs, higher API limits, Slack / Zoom. Enterprise: custom SLAs, self-host, flex compute. New accounts get free credits (FAQ).

Model API examples (per 1M tokens, input / cache / output)

• Kimi K3 — $3.00 input / $0.30 cache-input / $15.00 output per 1M on the table we copied (cover).
• Other rows (elephant-1, GLM-class, etc.) share that grid — copy the model you will call.

Dedicated compute (per minute)

• H100 80 GB — $0.10833/min (~$6.50/hr)
• A100 80 GB — $0.06667/min (~$4.00/hr)
• B200 180 GB — $0.16633/min (~$9.98/hr)
• L4 — $0.01414/min

You pay while a replica is deploying, scaling, or serving — not while it is scaled to zero.

Limitations / costs

• No public $0 hosted API. Trial credits expire; then you bill.
• Always-on H100 is not cheap. Two replicas × $6.50/hr is a monthly five-figure line.
• Basic vs Pro is support and quota, not a different GPU sticker.
• Not a Pod marketplace. For that, see Runpod or Vast.ai.

Alternatives

• Fireworks AI — hosted inference if you want a public token card and no Truss.
• Together AI — open-weight APIs if you will not deploy your own checkpoint.
• Modal — functions, not a model-serving control plane.

Official tutorial

We searched the official Baseten channel for a current getting-started / product-UI walkthrough. Recent uploads are briefings and partner talks, not a first-party “open the console” tutorial, so nothing is embedded. Use the written Quickstart plus Pricing.

Website

https://www.baseten.co

Repository

https://github.com/basetenlabs/truss

Source

Catalog row curated for AI Tools Nav. Cover is Baseten’s official Model APIs price table. H100 $0.10833/min. Kimi K3 $3.00 / $0.30 / $15.00 per 1M. Official channel: @baseten (no getting-started embed).

Last verified

2026-10-22 (Asia/Shanghai) editorial check.