Runcrate is a prepaid cloud where you rent a GPU with a shell, or call open models through one API, and both draw from the same balance.

Two products, one balance

You put money on the account first. A GPU machine and a model call spend that same balance. There is no monthly fee. Runcrate says the balance does not expire. This page did not launch a machine or call a model.

An instance is a machine with a named GPU, SSH, Jupyter, and VS Code in the browser. Ready images called crates, including Devbox, Finetune, and Inference, boot with the tools already installed. Runcrate says the image itself is not an extra charge on top of the machine.
The model API is OpenAI-compatible at api.runcrate.ai/v1. Their pricing page, checked 22 October 2026, says 223 models sit behind it. One listed rate is DeepSeek V3.2 at $0.273 per million input tokens and $0.399 per million output tokens. Image models are mostly priced per image, and video per second.
A storage volume keeps files after the GPU is deleted, then mounts on the next machine. The price is $0.03 per GB a month, from 10 GB to 10,000 GB, prepaid a week at a time. Delete it early and the unused days go back to the balance. Their docs say the files live on the instance disk at /workspace, and one volume attaches to one running instance at a time.
An agent can drive it from the terminal. The documented Claude Code setup is one command: claude mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcp. You then approve the login in the browser. There is no key to paste for that step.

The from-price is the cheapest offer, not the only one

Runcrate publishes a price feed at runcrate.ai/api/gpu-pricing and says it refreshes every five minutes. The copy below is that feed at 13:00 UTC on 22 October 2026. It listed 17 GPU types. The low number is the cheapest offer in stock. The same chip is often listed higher.

A4000: $0.20 to $1.04 an hour. That $0.20 floor is the cheapest line in the feed.
A6000: $0.65 to $2.51 an hour. An October 19 launch replay on the homepage used an A6000 at $0.65 an hour, which matches today's floor, not every A6000 they list.
H100 SXM: $5.17 to $7.79 an hour, and the feed only showed single-GPU offers.
H200: $4.24 an hour, flat in this snapshot. It is the only GPU the feed marks as billed by the minute. The others in the feed are not marked that way.
GAUDI2: $1.43 an hour, and only as an eight-GPU count.
Eight-GPU counts also appear as on-demand offers for H200, A100, L40, and a few others. A reserved quote is a separate path, for a month or longer, several nodes, or a cluster an engineer sizes.

The same pricing page has a card saying $20 buys five hours on one H100 SXM. At the feed's cheapest H100 rate, $5.17 an hour, $20 is under four hours. This page does not decide which of those two numbers the cashier uses. Treat the feed as the list of offers, and the card as a slogan that did not match it.

When the meter runs

Billing starts when the instance reports running, and stops when you delete it. Boot time is not billed.
Most GPUs bill by the second. Runcrate says a few bill in whole minutes, rounded up, and that rounding includes about 30 seconds of teardown after you delete. In the feed above, only the H200 was flagged as per-minute.
You add $20 to $10,000 by card. A launch needs $10, or one hour of that machine, whichever is more. With auto-recharge on, the minimum they state is $10.
Auto-recharge uses a floor you set from $20 to $5,000 and a refill target from $50 to $10,000. A declined card turns it off, and they email you.
Under $1 they try auto-recharge, email you, and wait 15 minutes before deleting running instances. If the balance reaches minus $10, they say the wait can be shorter. Time already used is still billed.

Their own timing and spend figures

These numbers are Runcrate's, from their site, not a Seailife test.

Across 449 launches from 19 September to 18 October 2026, they report a median of 4 minutes 14 seconds from the launch request to a running machine with an address. The fastest tenth were inside 2 minutes 55 seconds. The slowest tenth took longer than 6 minutes 50 seconds.
An October 19 replay shows an A6000 queued, running at 4 minutes 15 seconds, then deleted. They say the address and ids were hidden.
From the same 30 days of settled customer runs, they publish medians: eight H200s, 18 runs from 8 teams, a median of 1 hour 47 minutes and $62.60. One H100 SXM, 26 runs from 14 teams, a median of 25 minutes and $1.44. One L40S, 47 runs from 5 teams, a median of 3 hours 9 minutes and $3.06. One H200, 14 runs from 5 teams, a median of 18 minutes and $1.34. Each team weighs the same in their method. Boot time is left out because it is not billed.

Who says they built it

On 27 October 2025, a Hugging Face post signed Dryft, under the handle heydryft, says Runcrate started after a bad stretch renting GPUs on vast.ai while trying to build a voice agent in April 2023. The post says the aim was stock you can take without a contract or a reservation, plus a browser VS Code button. LinkedIn's company page lists Nirupam Bhowmick as founder. That LinkedIn page also talks about model counts and B200 access that do not match the site checked for this listing, so those claims are not repeated here. No public code repository turned up on the homepage.

Where to look

The product is at runcrate.ai. Live GPU offers are the JSON at /api/gpu-pricing. The billing rules are on the pricing page. The origin post is on Hugging Face. The image on this page is Runcrate's wordmark banner, a drawing of stacked cards, not a screenshot of the console and not a Seailife test.

Two products, one balance

An instance is a machine with a named GPU, SSH, Jupyter, and VS Code in the browser. Ready images called crates, including Devbox, Finetune, and Inference, boot with the tools already installed. Runcrate says the image itself is not an extra charge on top of the machine.

The model API is OpenAI-compatible at api.runcrate.ai/v1. Their pricing page, checked 22 October 2026, says 223 models sit behind it. One listed rate is DeepSeek V3.2 at $0.273 per million input tokens and $0.399 per million output tokens. Image models are mostly priced per image, and video per second.

A storage volume keeps files after the GPU is deleted, then mounts on the next machine. The price is $0.03 per GB a month, from 10 GB to 10,000 GB, prepaid a week at a time. Delete it early and the unused days go back to the balance. Their docs say the files live on the instance disk at /workspace, and one volume attaches to one running instance at a time.

An agent can drive it from the terminal. The documented Claude Code setup is one command: claude mcp add --transport http --scope user runcrate https://mcp.runcrate.ai/mcp. You then approve the login in the browser. There is no key to paste for that step.

The from-price is the cheapest offer, not the only one

A4000: $0.20 to $1.04 an hour. That $0.20 floor is the cheapest line in the feed.

A6000: $0.65 to $2.51 an hour. An October 19 launch replay on the homepage used an A6000 at $0.65 an hour, which matches today's floor, not every A6000 they list.

H100 SXM: $5.17 to $7.79 an hour, and the feed only showed single-GPU offers.

H200: $4.24 an hour, flat in this snapshot. It is the only GPU the feed marks as billed by the minute. The others in the feed are not marked that way.

GAUDI2: $1.43 an hour, and only as an eight-GPU count.

Eight-GPU counts also appear as on-demand offers for H200, A100, L40, and a few others. A reserved quote is a separate path, for a month or longer, several nodes, or a cluster an engineer sizes.

When the meter runs

Billing starts when the instance reports running, and stops when you delete it. Boot time is not billed.

Most GPUs bill by the second. Runcrate says a few bill in whole minutes, rounded up, and that rounding includes about 30 seconds of teardown after you delete. In the feed above, only the H200 was flagged as per-minute.

You add $20 to $10,000 by card. A launch needs $10, or one hour of that machine, whichever is more. With auto-recharge on, the minimum they state is $10.

Auto-recharge uses a floor you set from $20 to $5,000 and a refill target from $50 to $10,000. A declined card turns it off, and they email you.

Under $1 they try auto-recharge, email you, and wait 15 minutes before deleting running instances. If the balance reaches minus $10, they say the wait can be shorter. Time already used is still billed.

Their own timing and spend figures

These numbers are Runcrate's, from their site, not a Seailife test.

Across 449 launches from 19 September to 18 October 2026, they report a median of 4 minutes 14 seconds from the launch request to a running machine with an address. The fastest tenth were inside 2 minutes 55 seconds. The slowest tenth took longer than 6 minutes 50 seconds.

An October 19 replay shows an A6000 queued, running at 4 minutes 15 seconds, then deleted. They say the address and ids were hidden.

From the same 30 days of settled customer runs, they publish medians: eight H200s, 18 runs from 8 teams, a median of 1 hour 47 minutes and $62.60. One H100 SXM, 26 runs from 14 teams, a median of 25 minutes and $1.44. One L40S, 47 runs from 5 teams, a median of 3 hours 9 minutes and $3.06. One H200, 14 runs from 5 teams, a median of 18 minutes and $1.34. Each team weighs the same in their method. Boot time is left out because it is not billed.

Who says they built it

Runcrate

Two products, one balance

The from-price is the cheapest offer, not the only one

When the meter runs

Their own timing and spend figures

Who says they built it

Where to look

Comments

Runcrate

Two products, one balance

The from-price is the cheapest offer, not the only one

When the meter runs

Their own timing and spend figures

Who says they built it

Where to look

Comments