Baseten iconBaseten

Inference platform with dedicated GPU deployments billed per minute, per-token Model APIs for open models, training compute and agent sandboxes

$1.5B Series F (Jun 2026)
T4 to B200 Dedicated GPUs
2019 Founded

Overview

Baseten is an inference platform for teams that run AI models in production. It sells two main ways to serve models: dedicated deployments, where you run your own or open models on GPUs from T4 up to B200 and pay per minute, and Model APIs, where you call hosted open models such as GPT OSS 120B, DeepSeek, GLM and Kimi and pay per token, with cheaper cached input. The same GPU rates apply to training jobs, and newer product lines add serverless "Loops" endpoints (prefill, sample and train) for reinforcement learning work and CPU sandboxes for agents. Its public customer list includes Abridge, Canva, Cognition, Coinbase, Notion, Ramp, Vercel, Writer and Zed. The company is based in San Francisco, is led by co-founder and CEO Tuhin Srivastava, and raised a $300M Series E in January 2026 and a $1.5B Series F in June 2026, for over $2B raised in total. Compare it with Together AI, Fireworks AI and Modal.

The Verdict

Who Should Use Baseten?

Best For

  • Product teams serving custom or fine-tuned models at scale
  • Teams that want both per-token APIs and dedicated GPUs from one vendor
  • Companies that need reserved capacity and custom terms at high volume

Not Ideal For

  • Hobby projects looking for a free tier (none is listed)
  • Non-technical teams without ML or platform engineers

What's Great

  • Public per-minute GPU rates and per-token model prices
  • Wide GPU range, from T4 to B200, with H100 MIG slices
  • Large named customer base in production AI products

Watch Out For

  • No free credits shown on the pricing page
  • Volume pricing and reserved compute are contact-sales only
  • Few public user reviews: G2 has no reviews yet

Pricing

View all features & details

Key Features

  • Dedicated deployments on T4, L4, A10G, A100, H100 and B200 GPUs
  • Model APIs for hosted open models, priced per token with cached input
  • Training compute at the same GPU rates
  • Serverless Loops endpoints for prefill, sample and train
  • Sandboxes at $0.0655 per vCPU-hour and $0.0211 per GB-RAM-hour
  • CPU-only instances

Company

  • CEO and co-founder Tuhin Srivastava
  • Series E: $300M at a $5B valuation (Jan 2026)
  • Series F: $1.5B led by Altimeter, Conviction and Spark (Jun 2026)

How It Compares

Feature Baseten Together AI Fireworks AI
Per-token open-model APIs Yes Yes Yes
Dedicated GPUs Per minute, T4 to B200 Yes Yes
Training GPU rates, plus Loops for RL Fine-tuning service Fine-tuning service

Sources: Baseten · Pricing (fetched Nov 2026) · Customers · Series F coverage (Jun 2026)

Is this your tool? Add a badge