Pricing at a glance
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Decision point | What Runcrate states |
|---|---|
How you pay | One prepaid balance, added by card before use, pays for GPUs, models and storage. No monthly fee, and the balance does not expire. |
Top-ups | $20 to $10,000 at a time. Auto-recharge takes a threshold of $20 to $5,000 and a target of $50 to $10,000; a declined card turns it off. |
Where prices come from | Live rates: the GPU table shows the lowest live rate for each GPU, also served as JSON at /api/gpu-pricing with no key, refreshed every 5 minutes. A sold-out GPU stays listed for 30 days at 0 available with its last price. |
Which way to buy
Pay as you go suits bursty work: GPU instances bill per GPU-hour, by the second on most GPUs and by the minute on a few, and booting is never billed.
A prepaid machine suits steady use of one GPU: prepay it for 1 to 12 months at launch, at a discount for paying ahead, billed by Runcrate.
Self-serve clusters suit independent jobs across machines: one hourly price for the whole cluster, every machine starts or none does, and a cluster that could not start costs nothing. They run on Ethernet, which Runcrate says is not for training one model across machines.
Reserved clusters are for a month or longer: Runcrate verifies your company and need, takes your spec to GPU owners that fit and shows their offers side by side, and you sign with the owner you pick.
Inference is per 1M tokens, input and output priced separately; most image models bill per image and video per second. Five open models are free with a GitHub account, then half price on shared capacity and list price when it is busy.
Rates on November 1, 2026
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| GPU | Memory | From, per GPU-hour |
|---|---|---|
B300 | 288GB | $8.21 |
B200 | 180GB | $7.67 |
H200 | 141GB | $4.63, billed by the minute |
H100 | 80GB | $2.49 |
A100 80GB | 80GB | $1.73 |
L40S | 48GB | $0.54 |
RTX 4090 | 24GB | $0.57 |
V100 32GB | 32GB | $0.36 |
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Model | Input | Output | Cached input |
|---|---|---|---|
Kimi K2.6 | $0.788 | $3.68 | $0.158 |
Qwen3 Coder 480B A35B | $0.315 | $1.05 | $0.105 |
Llama 4 Scout 17B 16E | $0.105 | $0.315 | No cache discount |
How the bill adds up
A GPU's meter starts when the instance is running and stops when you delete it. Each model call is priced by its tokens, images or seconds of output and drawn from the same balance. A storage volume costs $0.03 per GB a month, prepaid a week at a time while it exists, and unused days come back to the balance if you delete it early.
Runcrate's worked figures
What $20 buys: 8 hours on one H100, 11 hours on one A100 80GB or 37 hours on one L40S; or, on DeepSeek V3.2, 73M input tokens at $0.273 per 1M or 50M output tokens at $0.399 per 1M.
Two real charges, October 19, 2026: an A6000 booted in 4 min 12 s, not billed, then ran 7.9 s at $0.65/hr for $0.00143; a DeepSeek-V3 call with 18 tokens in and 63 out cost $0.0000649. Total: $0.00149.
Runcrate's H100 ledger: across 26 runs from 14 teams, September 19 to October 18, 2026, the median 1 × H100 run lasted 25 minutes and cost $1.44; the middle half cost $0.40 to $8.94.
Storage for a month: 100 GB is $3.00, 1 TB $30.00 and 10 TB $300.00.
Costs that are easy to miss
Teardown on per-minute GPUs. A few GPUs bill in whole minutes, rounded up, including about 30 seconds of teardown after you delete.
Instances you forget. An instance bills until you delete it; give it a maximum runtime (5 minutes to 168 hours) or an idle stop (5 minutes to 24 hours), which terminates it.
Balance needed to launch. A launch needs $10 or an hour of the machine, whichever is more, or just $10 with auto-recharge on.
A low balance deletes running work. Under $1.00, Runcrate tries auto-recharge, emails you and waits 15 minutes before deleting running instances, or sooner once the balance reaches -$10.00.
Our take
Runcrate's pricing suits work that ends and can be restarted: the meter stops when you delete the instance, and a volume keeps your files for the next one. For a service that must stay up, the prepaid balance is the main risk, so turn auto-recharge on before you rely on it. For many short jobs, prefer GPUs that bill by the second, because per-minute GPUs round each run up and add about 30 seconds of teardown.