Replicate vs Together AI vs Fireworks AI: what running AI models costs

Updated 20 November 2026

First published 2026-11-20 · Tool Comparison

Checked 20 November 2026. Every price below was read on the vendor’s own page that day and is quoted as printed. A vendor with no printed price says so; we did not fill the gap.

TL;DR: Nov 2026: H100 per hour as printed: Replicate $5.49, Together $5.49 for dedicated inference, Fireworks $8.00. All prices are in dollars. Checked 20 November 2026.

At a glance

ToolWho it suitsOne figure, as printed
ReplicateA team that runs public models and pays by the second of run time, or rents dedicated hardware for its own model.Billing basis: “The price-per-second varies according to the hardware in use.”
Together AIA team that wants per-token rates for open models and rented GPUs, with a clusters table that shows three ways to pay.Dedicated inference, H100: “NVIDIA HGX H100 | $5.49 | Contact sales”
Fireworks AIA team that wants a per-token price for open models and a GPU rate card billed by the second.Serverless: “Serverless: Pay per token · From $0.05 / 1M input tokens.”

Replicate

A team that runs public models and pays by the second of run time, or rents dedicated hardware for its own model.

  • Billing basis: “The price-per-second varies according to the hardware in use.” source, read 20 November 2026.
  • Nvidia H100: “Nvidia H100 GPU (gpu-h100) $ 0.001525 /sec, $ 5.49 /hr” source, read 20 November 2026.
  • Nvidia A100 80GB: “Nvidia A100 (80GB) GPU (gpu-a100-large) $ 0.001400 /sec, $ 5.04 /hr” source, read 20 November 2026.
  • Nvidia T4: “Nvidia T4 GPU (gpu-t4) $ 0.000225 /sec, $ 0.81 /hr” source, read 20 November 2026.

Not read. The page lists per-model prices too (for example images and video); we quote none of them. The helper's copy of this page is in table form, not word for word.

Together AI

A team that wants per-token rates for open models and rented GPUs, with a clusters table that shows three ways to pay.

  • Dedicated inference, H100: “NVIDIA HGX H100 | $5.49 | Contact sales” source, read 20 November 2026.
  • GPU clusters, H100: “NVIDIA HGX H100 | $1.99 | $3.99 | $3.69 | $3.45 | Request Capacity” source, read 20 November 2026.
  • Basis of the cluster prices: “All prices are per GPU per hour.” source, read 20 November 2026.
  • Kimi K3, per 1M tokens: “Kimi K3 | $3.00 $0.30 (cached) | $15.00” source, read 20 November 2026.
  • DeepSeek V4.1 Flash, per 1M tokens: “DeepSeek V4.1 Flash 1017 | $0.30 $0.006 (cached) | $1.20” source, read 20 November 2026.

Not read. The helper's copy of this page is in table form, not word for word. Its column headings for the cluster table are “Preemptible”, “On-demand” and reserved periods of “3-30 days”, “31-90 days” and “91+ days”; we quote the row as returned and do not pair each figure to a heading beyond that. Live cluster prices carried a “Checking…” label.

Fireworks AI

A team that wants a per-token price for open models and a GPU rate card billed by the second.

  • Serverless: “Serverless: Pay per token · From $0.05 / 1M input tokens.” source, read 20 November 2026.
  • Kimi K3, per 1M tokens: “Kimi K3: $3.00 | $0.30 | $15.00 | $3.75 / $0.375 / $18.75” source, read 20 November 2026.
  • DeepSeek V4.1 Flash, per 1M tokens: “DeepSeek V4.1 Flash: $0.30 | $0.006 | $1.20 | $0.375 / $0.0075 / $1.50” source, read 20 November 2026.
  • H100 80 GB: “H100 80 GB | $0.134 | $8.00” source, read 20 November 2026.
  • Billing basis: “On Demand bills per GPU-second.” source, read 20 November 2026.

Not read. The helper's copy of this page is in table form, not word for word. The columns of the GPU table are “GPU”, “Per minute” and “Per hour”; the token table's columns are “Model”, “Input”, “Cached input”, “Output” and “Priority”.

Open score and India fit

Our open score gives a tool up to 60 points from signals we can check: rupee pricing, GST invoice, UPI, data in India, a public price, a free plan and the number of alternatives. “Not recorded” means we have not found a vendor statement on that point, not that the answer is no. How the score is worked out.

ToolOpen scorePrices in ₹GST invoiceUPIData in India
ReplicateNo open score yetNot recordedNot recordedNot recordedNot recorded
Together AINo open score yetNot recordedNot recordedNot recordedNot recorded
Fireworks AINo open score yetNot recordedNot recordedNot recordedNot recorded

Before you decide

  • Replicate's H100 row is hardware for a model you run; Together's $5.49 is dedicated inference and its $3.99 is an on-demand cluster price; Fireworks' $8.00 is its GPU rate card. These are different products at one vendor each, so the three H100 figures are not a ranking.
  • Two of the three print the same two models (Kimi K3 and DeepSeek V4.1 Flash) at the same input and output rates in their token tables, as returned to us; check the live page before you rely on either.
  • All three print dollars; none printed a rupee price. Where a page prints dollars, a card payment in dollars may carry a foreign-exchange fee from the card issuer; the pages do not say.

What AI assistants name

Our AI Visibility Index asks assistants about 30 software categories each month. AI inference hosting is not one of them yet, so there is no “who AI names” line for it on this page.

Deals and affiliate links

This page carries no deal or affiliate links. Nobody paid to be on it and the order of the tools is not a ranking.

Check what AI says about your product

Buyers now ask an assistant which AI inference host to pick. If you sell one, see what it says about you: three buyer questions, about 15 seconds, no sign-up.

Run the free AI check for ai inference hosting →

The Daily Brief — a daily update across 12 industries

One actionable growth breakdown every morning, across 12 industries — with an audio version in English and 10 Indian languages. No fluff, just hard product teardowns and India benchmarks.

or

Not sure which one fits your case? Ask Kriyā with the options in the question.

Ask Kriyā →