L4 price per hour
Ada inference card, 24GB.
The cheapest firm L4 in this print is $0.44 per hour at Runpod, L4 24GB, community. Captured 17 Oct 2026, 00:18 UTC.
Questions
What is the cheapest L4 per hour?
As of Oct 17, 2026 the cheapest firm L4 is $0.44/GPU-hr (Runpod Community Cloud, L4 24GB); Koyeb on-demand is $0.70.
What is the cheapest spot L4?
No spot or interruptible L4 quote was published in the Oct 17, 2026 print.
How did L4 prices move this week?
L4 at Runpod (L4 24GB) is up 20.4% versus Oct 9, 2026, from $0.49 to $0.59/GPU-hr.
On-demand vs spot: what's the difference?
On-demand, Runpod Community Cloud, and a published starting price are firm L4 quotes: the rate printed for capacity without a preemption warning. Spot and interruptible rates can be lower, and the host can take the GPU back. This page keeps them out of the cheapest firm line. A MIG slice is a fraction of a card and is labeled as such. Reserved terms are listed and do not set the headline.
Price per hour by provider
| Provider | Board | Type | $/GPU-hr | Week | Source |
|---|---|---|---|---|---|
| Runpod L4 24GB Form factor not printed Community | Community | $0.44 | — | Page | |
| Runpod L4 24GB Form factor not printed On-demand | On-demand | $0.59 | +20.4% | Page | |
| Koyeb L4 24GB Form factor not printed On-demand | On-demand | $0.70 | — | Page |
Week compares the same provider, board, price type, and GPU count with a print from about seven days earlier. A dash means that pair was not in the earlier print.
Firm quotes
History of the low
Use this chart
Drop the card into a post, or use the badge. Both link back here. The price table itself is CC BY 4.0.
<iframe src="https://gpu-price-index.pages.dev/embed/l4" width="320" height="120" loading="lazy" title="L4 price per hour"></iframe><a href="https://gpu-price-index.pages.dev/gpu/l4"><img alt="L4 from $0.44/hr · GPU Price Index" src="https://gpu-price-index.pages.dev/badge/l4.svg"></a>About this L4 page
“L4 price per hour” here means the published USD rate for one GPU, or a node price divided by the GPU count when the provider only sells a multi-GPU machine. The usual memory on this board is 24 GB; a row says otherwise when the source prints a different figure.Booking still happens on the provider's site. Referral links on provider pages are placeholders and are not paid placements.