---
title: "Announcing StarSling GPU Runners"
description: "GitHub Actions runners backed by NVIDIA GPUs: RTX PRO 6000, RTX 5090 and RTX 4090, with H100 coming. One line in runs-on. Contact us to get started."
date: 2026-09-10
url: https://starsling.dev/blog/announcing-starsling-gpu-runners
canonicalUrl: https://starsling.dev/blog/announcing-starsling-gpu-runners
---

# Announcing StarSling GPU Runners

<figure>
  [Image: A StarSling card headed Announcing StarSling GPU Runners. Beneath it, four GPU packages sit on one bus, labelled RTX PRO 6000, RTX 5090, RTX 4090 and H100, the last marked coming soon, above the runner label runs-on: starsling-ubuntu-24.04-gpu.]
</figure>

Today we're announcing [**StarSling GPU Runners**](/products/gpu-runners): GitHub Actions runners backed by NVIDIA GPUs. [Contact us to get started](https://tally.so/r/Y56E8v).

StarSling is AI-native CI for GitHub Actions: fast runners with agents that read your real runs and open pull requests that make them faster. GPU Runners are that same platform but allow for GPU specific code to be exercised as part of a GitHub Actions workflow.

## Your production GPUs should run your product

A year ago very few teams needed a GPU for CI. Now a lot of them serve their own models, and that work only gets tested when the model runs inside the job: self-hosted inference, serving code, quantization and CUDA changes, and the latency, throughput and memory numbers that come with them. Those checks are becoming ordinary tests, which means they want to run on every pull request like ordinary tests do.

The capacity those teams reach for is the capacity already serving production. A test suite borrows four GPUs from inference for twenty minutes, the pool it borrowed from gets thinner, and the tradeoff lands on whoever is on call that week. Then an agent opens another PR and it happens again.

GPU Runners give those jobs their own hardware, inside the GitHub Actions workflow you already have. The GPU is dedicated to the job for as long as the job runs, and you pay for it by the minute. A twenty minute suite gets twenty uninterrupted minutes on the card, which is the part that matters in CI: a GPU reclaimed halfway through a test run is a red check on a pull request that had nothing wrong with it.

## The lineup

| GPU | Best for | Per minute |
| --- | --- | --- |
| NVIDIA RTX PRO 6000 (default) | Large-model and memory-heavy CI workloads | $0.05922 |
| NVIDIA RTX 5090 | Fast inference tests and model evals | $0.03022 |
| NVIDIA RTX 4090 | Everyday GPU tests at the lowest cost | $0.02522 |
| NVIDIA H100 | Maximum throughput for large-model CI workloads | Coming soon |

The three RTX GPUs come on a 4 vCPU host with 16 GB of memory and 100 GB of disk. For the latest labels and pricing, see [the StarSling docs](https://docs.starsling.dev/runners/compute-sizing#gpu-specifications).

## One line in `runs-on:`

The bare `starsling-ubuntu-24.04-gpu` label gives you the RTX PRO 6000. To pick a different GPU, add the selector `gpus=<sku>:<count>`; runners are single-GPU today, so the count is 1. The accepted SKUs and the full selector grammar are in [the label reference](https://docs.starsling.dev/configuration/label-reference#gpu-labels).

That's the migration. Same workflow file, same actions, same secrets, same review process. Every job runs in its own single-use, hardware-isolated microVM that is destroyed after the run, and billing works the way it does on a CPU runner: job start to finish, rounded up to the nearest minute, queue time free and concurrency unlimited.

## Getting access

[Contact us to get started](https://tally.so/r/Y56E8v) and we'll email you when your org is up: we're granting access on a rolling basis based on configuration and load preferences, so be sure to fill that out in the form.

If your team is new to StarSling, [install the StarSling GitHub App](https://github.com/apps/starslingdev) and point a job at [StarSling Runners](/products/runners) today. Your first month includes 2,000 free minutes. When your GPU access lands, the workflow change is one more line.

The full lineup, prices and host specs live on the [StarSling GPU Runners page](/products/gpu-runners).

Happy Slinging 

Yonas