Skip to content
runcrate
Console

GPUs and inference for your agents.

Your agent launches a GPU with a real shell in about four minutes, or calls 223 models through one OpenAI-compatible API. One balance pays for both.

Get startedConnect your AI agent

16 GPU models from $0.51/hr, billed by the second on most.

runcrate instances create --name demo --gpu A6000; runcrate instances status demo; runcrate ssh demo -- nvidia-smi; runcrate instances delete demo; runcrate ps
Outputinstances create
4:23running in 4:15$0.65/hrA6000
  1. 0:02Instance queued: demo (A6000)
  2. 0:04deploying
  3. 4:15running
  4. 4:20NVIDIA RTX A6000 · 49140MiB
  5. 4:22Instance terminated.
  6. 4:23No instances found.

Replayed from a real launch on October 19, 2026, the boot sped up.

Teams at these organizations use Runcrate

What you can build

Train, serve, render and transcribe. Real jobs from the docs, ready for your agent.

Examples

Hand it to your agent. Or read exactly what it runs.

Docs

Deploy, manage, and terminate GPU instances from the terminal.

Start an A6000, run nvidia-smi on it, then shut it down. Runs runcrate instances types --gpu A6000. It returns: ID GPU COUNT CPU RAM STORAGE REGION $/HR goofy-boyd-a6000-fa6f A6000 1 28 cores 58 GB 100 GB montreal-canada-2 $0.65 crazy-yalow-a6000-02b3 A6000 1 6 cores 48 GB 256 GB wichita-usa-1 $0.74. Runs runcrate instances create --name demo --gpu A6000. It returns: Instance queued: demo (A6000) Cost: $0.65/hr. Runs runcrate instances status demo. It returns: Status: ● running. Runs runcrate ssh demo -- nvidia-smi. It returns: | NVIDIA-SMI 570.195.03 Driver Version: 570.195.03 CUDA Version: 12.8 | | 0 NVIDIA RTX A6000 On | 00000000:00:06.0 Off | Off | | 30% 54C P8 13W / 300W | 1MiB / 49140MiB | 0% Default |. Runs runcrate instances delete demo. It returns: Terminating instance demo... Instance terminated..

NoteA real launch on October 19, 2026, at that day's prices. Ids and the address are hidden.

Products

Five products, one balance. Use one, or all of them on the same account.

POST /api/v1/instances
449 launches · Sep 19 to Oct 18, 2026

Measured

Launch to running in about four minutes. The wait is not billed.

The time from a launch request to the instance running with an address, across every launch on 16 GPU models in thirty days. Billing starts when the instance is running.

Median, launch to running

4:14

The fastest tenth were running within 2:55, the slowest tenth took longer than 6:50.

  1. All GPUs4:14
  2. RTX PRO 60002:59
  3. H2003:09
  4. RTX 6000 Ada3:12
  5. A60004:02
  6. L40S4:12
  7. H100 SXM4:14
  8. RTX 40904:34
  9. RTX 50904:40
  10. L405:27
  11. A100 80GB5:35

Median minutes:seconds per GPU model with at least 20 launches in the window, measured October 19, 2026.

Get started

A GPU from your agent in minutes. Inside the tools you already use.

  1. 01

    Connect

    One command adds Runcrate to Claude Code, Codex, Cursor or your terminal.

  2. 02

    Sign in

    Your browser opens and you approve. No key to paste.

  3. 03

    Ask

    “Launch an H100 with the Finetune crate and run nvidia-smi.” Billing starts when the instance is running.

Every tool your agent gets
terminalbash
claude mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcp

Then run /mcp in Claude Code and pick runcrate: your browser opens to sign in. No key to paste.

Why it works

You pay for the time the GPU runs. Not for the boot, and never for a whole hour you did not use.

Billing starts when the instance reports running. Most GPUs then bill by the second; a few bill in whole minutes, rounded up, including about 30 seconds of teardown after you delete. Keep your files on a volume and you can delete the GPU between jobs without losing work.

  • Nothing billed while it boots
  • By the second on most GPUs
  • Volumes keep files between instances
See prices and how billing works

Pricing

Dollars per GPU-hour. Per token for inference.

One prepaid balance pays for GPUs, models and storage. These prices are live and are what you are billed.

  • GPU instances

    For training, fine-tuning and serving

    from$0.51/ GPU-hour

    By the second on most GPUs, a few by the minute

    Booting is never billed

    Launch a GPU

    $20 buys one of

    • 4 hon one H100 SXM
    • 11 hon one A100 80GB
    • 17 hon one L40S

    Includes

    • SSH, Jupyter and VS Code in the browser
    • 16 GPU models on offer
    • Ready images for fine-tuning and serving

  • Inference

    For apps and agents calling models

    $0.273/ 1M input tokens

    $0.399 per 1M output, on DeepSeek V3.2

    Most images per image, video per second

    Get an API key

    $20 buys one of

    • 73Minput tokens on DeepSeek V3.2
    • 50Moutput tokens on DeepSeek V3.2
    • 761images from FLUX.1 Dev

    Includes

    • 223 models behind one API
    • OpenAI-compatible, so your client already works
    • No monthly fee

  • Storage volumes

    For datasets, weights and checkpoints

    $0.03/ GB a month

    Prepaid a week at a time while it exists

    Unused days refunded on delete

    Create a volume

    A month of

    • $3.00100 GB
    • $30.001 TB
    • $300.0010 TB

    Includes

    • Kept when you delete the instance
    • Mounted on the next instance you launch
    • 10 GB to 10,000 GB a volume

Reserved clusters. For 8 GPUs or more, several nodes, or a month or longer: an engineer sizes the cluster with you and prices it for the term.

Talk to an engineer
See every GPU and model price