Baseten vs Fireworks AI vs Together AI: pricing compared

Baseten vs Fireworks comes down to how each company bills dedicated GPUs, what it publishes for serverless, and what it offers for fine-tuning. This comparison covers Baseten, Fireworks AI and Together AI, using their own pricing pages and docs as read on October 31, 2026. Baseten and Together publish the same per-token price for the models they share, Together is cheapest per hour on dedicated H100 and B200, and Fireworks does not list per-token chat prices on its pricing page.

All three sell the same core thing: hosted open models behind an API, plus dedicated GPUs for production traffic and training for custom models. The differences are in the details of pricing and packaging.

This comparison is based on the products' docs and published results, not hands-on testing. Prices change, so check each page before you commit. Everything below is as of October 31, 2026.

Products side by side

Sources: Baseten docs, Fireworks docs, Together docs.

Serverless pricing per million tokens

Baseten and Together list the same price for every model they have in common that we checked. Fireworks' pricing page links out to its docs for chat model rates and shows only embeddings on the page itself, so a like-for-like row for Fireworks is not possible from that page.

The cached-input rates differ in places. For GLM-5.3, Baseten lists $0.14 and Together lists $0.26. For DeepSeek V4.1 Flash, Baseten lists $0.007 and Together $0.006. Fireworks' page shows embeddings at $0.008 per million input tokens for models up to 150M parameters and $0.016 for 150M to 350M.

Dedicated GPU pricing

Baseten bills by the minute and only for compute time, not idle time, per its pricing page. Fireworks lists per-minute and per-hour rates. Together lists per-hour rates. The Baseten hourly figures below are our arithmetic, per-minute price times 60.

Fireworks adds a 1.5x premium for region-restricted deployments. Baseten says compute in other regions is available by contacting sales, and discounts can be negotiated on Pro and Enterprise. Together also lists separate GPU cluster rates, for example H100 at $3.99 an hour on demand and $1.99 preemptible.

The units differ across pages, and the products may not be configured the same way, so treat these as list prices, not total cost for a workload.

Fine-tuning pricing

The three companies price training in different ways, which makes a direct comparison hard.

• Baseten lists no separate fine-tuning price. Training uses the same per-minute GPU rates as deployments, for example $0.10833 a minute on an H100, and is included in the Basic plan.
• Fireworks charges per million training tokens. LoRA supervised fine-tuning costs $0.50 for models up to 16B parameters and $3.00 for 16.1B to 80B. Full-parameter SFT costs $1.00 and $6.00 for the same sizes. It also lists a serverless training API, and says Muse Glimmer 30B SFT is free until 11/30.
• Together charges per million tokens by model. Llama 3.3 70B costs $2.03 for SFT and $5.08 for DPO, with a $4.00 minimum. Llama 3.1 8B costs $0.34 and $0.84.

What each one is best for, going by the docs

Baseten's pitch is dedicated deployments with autoscaling, plus its own training tools. If you run custom or private models, that is the closest fit. Fireworks gives you the most training options on its pricing page, with LoRA and full-parameter, SFT and DPO, and a serverless training API. Together has the widest menu on one page: serverless text, image, audio and video models, dedicated endpoints, fine-tuning and raw GPU clusters.

None of this tells you which is fastest or most reliable. The pages above do not publish latency or throughput benchmarks, and we did not test any of the three.

What we could not confirm

We did not find per-token chat prices on Fireworks' pricing page, so we left them out. Baseten's page does not state the amount of free credits for new accounts, and neither Fireworks nor Together documents a free credit amount there. We also could not open the Techmeme headline about a reported Baseten funding round, and a search did not turn up the Bloomberg story, so this article does not report on it.

Before choosing, price your own workload: the model you need, the tokens per month and whether traffic is steady enough to justify a dedicated GPU.