Baseten pricing, plans and limits

Baseten is a production AI inference platform for ML teams that serve custom or open-weight models. The Basic plan has no monthly fee and GPUs bill by the minute, so a deployment that scales to zero costs nothing idle. Model APIs bill per token. Baseten is SOC 2 Type II and HIPAA compliant, and it bought Blaxel in September 2026.

Baseten is an AI inference platform founded in 2019 that runs custom and open-weight models as autoscaling endpoints, on dedicated GPUs or through per-token Model APIs. Its multi-cloud layer routes replicas across many providers, and the company says it has raised more than $2B. It is built for production ML teams, and it states HIPAA compliance for healthcare buyers.

About Baseten

Baseten is a San Francisco AI inference platform founded in 2019 by Tuhin Srivastava, Philip Howes, Amir Haghighat and Pankaj Gupta. Engineers package a model with the open-source Truss framework, push it, and get an autoscaling HTTPS endpoint while Baseten handles GPU provisioning, weight caching and observability. On its own site Baseten lists Cursor, Notion, Clay, Abridge, HeyGen, Descript, Harvey and OpenEvidence among its customers.

The platform now has three product lines: dedicated deployments (your own model on a reserved-by-the-minute GPU), Model APIs (per-token access to open-weight models such as Kimi K3, GLM 5.3, DeepSeek V4 Flash and DeepSeek V4 Pro 0813, plus Meta's Muse Spark Open and two small decision models, elephant-1 and mammoth-1, from Baseten's own Base Labs group that return probabilities for routing and classification instead of text), and Training for fine-tuning on the same infrastructure. Its multi-cloud capacity management layer treats GPUs from over twenty clouds and regions as one pool, and Baseten states 99.99% uptime for it. Weight caching through the Baseten Delivery Network keeps cold starts short after scale-to-zero. On 2026-09-11 Baseten announced that it had acquired Blaxel, a sandbox and agent-storage company, to add a place where long-running agents execute next to the models they call.

Funding has moved quickly. Bloomberg reported a $300M round at a $5B valuation in January 2026 led by IVP and CapitalG with Nvidia (Nvidia also invests), and Baseten's own llms.txt states a $1.5B Series F in June 2026 at a $13 billion valuation. Baseten's own Blaxel post says it has raised more than $2B. Those are company-level figures and say nothing about the price you will pay for a GPU-minute.

Pricing is usage-based: the Basic plan has no monthly fee, dedicated GPUs bill by the minute across seven instance types, and Pro and Enterprise are quoted through sales. The exact rates sit in the cost section of this page.

Where a rival wins: a team that only wants per-token access to popular open models is often simpler to serve on Together AI or Fireworks AI, and a team that wants a code-first Python function runtime will find Modal lighter. Teams that want bare GPUs at a flat hourly rate fit better on Crusoe or Atlas Cloud. Pipeshift aims at the same real-time-agent inference niche. Baseten is the stronger pick when you serve your own or fine-tuned models, need HIPAA and SOC 2 Type II, and want multi-cloud failover handled for you. Compare the wider field in Google Vertex AI if you already run on Google Cloud.

Pricing

The Basic plan costs $0 per month and bills pay-as-you-go by the minute for dedicated deployments, Model APIs and Training. Baseten lists per-minute rates, so the hourly figures here are those rates times 60 (T4 $0.63, L4 $0.85, A10G $1.21, A100 $4.00, H100 MIG $3.75, H100 $6.50, B200 $9.98). Model APIs bill per 1M tokens: text models run from $0.10 input and $0.50 output on GPT OSS 120B to $3.00 input and $15.00 output on Kimi K3, with cached input at a fraction of the input rate on most models, and the Base Labs decision models elephant-1 ($0.035) and mammoth-1 ($0.10) bill input tokens only.

Pro (priority GPU access, higher rate limits, Slack and Zoom support) and Enterprise (custom SLAs, self-hosted and hybrid deployments, data residency) are quoted through sales, and both carry volume discounts. Prices checked on the vendor page on 2026-10-29. Bandwidth is not listed as a separate line item.

Key Features

• Truss Open-Source Packaging: Truss is an open-source Python framework that packages a model into a production server container, so teams deploy PyTorch, TensorRT-LLM or vLLM models without writing Docker files.
• Multi-Cloud Capacity Management: MCM routes replicas across 20+ clouds and regions to work around capacity limits and outages, and Baseten states 99.99% uptime for it.
• Baseten Delivery Network: BDN mirrors model weights across node storage, an in-cluster cache and a mirrored origin so cold starts stay short after scale-to-zero.
• Model APIs: Per-token access to open-weight models including GLM 5.3, Kimi K3 and DeepSeek V4, with cached-input rates on most of them.
• Per-Minute Dedicated GPUs: Dedicated deployments bill by the minute across seven GPU types, with no charge for idle scaled-to-zero time.
• Training: Training runs fine-tuning jobs on the same GPU fleet and at the same per-minute rates as inference, so a model does not change platforms between tuning and serving.
• Blaxel Agent Infrastructure: Baseten has acquired Blaxel, adding microVM sandboxes, persistent storage and networking for long-running agents.

Strengths

• Cold starts stay short after scale-to-zero because model weights are cached close to the GPU, which matters most for large models that would otherwise reload for many seconds.
• SOC 2 Type II and HIPAA are stated on the pricing page and GDPR on the multi-cloud page, so healthcare and regulated buyers can start without a separate compliance project.
• Multi-cloud routing across many providers reduces exposure to a single provider running out of GPUs.
• Per-minute billing on seven GPU types makes small and bursty workloads cheaper to test than hourly-minimum rentals.
• Truss is open source, so packaging can be tested locally before any compute is paid for.

Weaknesses

• Pro and Enterprise prices are not public, so teams needing reserved capacity or self-hosting must go through sales.
• Per-plan vCPU and RAM for GPU instances are not listed on the pricing page, which makes right-sizing harder than on AWS or Google Cloud.
• The workflow assumes Python and a CLI, so non-technical operators cannot deploy models without an engineer.
• The Blaxel acquisition is only weeks old and the combined agent product has no published pricing yet.