Pricing · No seats · No platform fee
Billed by the token.Or flat, by the month.Never by the seat.
Two shapes of bill, because there are two things you might want from us. Send prompts and pay $0.04 per million input tokens on 4 hosted models (2 uncensored, 1M context on Qwen 3.8 Flash-Next Uncensored) — or pay a flat $20/mo and send as many messages as you like. New accounts start with $1.00 in credit and no card.
- Tokens
- $0.04
- per 1M input, cheapest of 4 catalogue models
- Shared capacity. You pay for what the model read and wrote, nothing for the card underneath it.
- GPU-hours
- $0.14/hr
- CMP-170HX spot, 9 cards on the menu
- Dedicated capacity. Billed on elapsed time and stopped the moment the deployment is, in either direction.
- Messages
- Unmetered
- flat plans from $20/mo
- A plan buys live context, not a message allowance. Nothing is counted against you on the way out.
Flat plans
One price a month. Messages are never metered.
A plan buys a live-context budget — how much of a model’s attention you can hold open at once — not a number of messages. Same $3.33 per GB of it on every tier.
Personal
$20per month
- Live context
- 6 GB
- Longest request
- 262k tokens
- In flight at once
- 16
- Private models
- 1
- One private model, always on
- Unlimited messages — never metered
- ~187,500 tokens of live context, split however you like
- 5 requests at 32k, or one at 187,500 — same budget
- Prompts never stored, never trained on
Pro
Most popular$100per month
- Live context
- 30 GB
- Longest request
- 262k tokens
- In flight at once
- 64
- Private models
- 3
- Up to 3 private models
- Unlimited messages — never metered
- ~937,500 tokens of live context
- Dozens of concurrent agent calls
- Run your own fine-tunes
- Prompts never stored, never trained on
Max
$200per month
- Live context
- 60 GB
- Longest request
- 262k tokens
- In flight at once
- 256
- Private models
- 8
- Up to 8 private models
- Unlimited messages — never metered
- ~1,875,000 tokens of live context
- Built for agent fleets — hundreds in flight
- Priority capacity during surges
- Run your own fine-tunes
- Prompts never stored, never trained on
Plans cover the models whose economics clear a flat rate; anything heavier is available per token or on a dedicated deployment below. Prompts on a plan are never stored and never used for training.
Per token · Serverless
Every rate we charge, in one table
4 models on shared capacity, 2 of them uncensored, all behind one OpenAI-compatible endpoint.
| Model | Input / 1M | Cached input / 1M | Output / 1M | Context | Decode | License |
|---|---|---|---|---|---|---|
| GPT-OSS 20BOpenAI · 21B MoE · MXFP4"gpt-oss-20b" | $0.04 | $0.01 | $0.16 | 131k | 260 tok/s | Apache 2.0 |
| Qwen 3.6 35B-A3BAlibaba · 35B-A3B MoE · FP8"qwen3.6-35b-a3b" | $0.10 | $0.02 | $0.40 | 262k | 220 tok/s | Apache 2.0 |
| Qwen 3.8 27B UncensoredUncensoredMoton · 27B · FP8"qwen3.8-27b-uncensored" | $0.25 | $0.05 | $0.75 | 262k | 74 tok/s | Apache 2.0 |
| Qwen 3.8 Flash-Next UncensoredUncensoredMoton · Flash MoE · FP8"qwen3.8-flash-next-uncensored" | $0.30 | $0.06 | $1.20 | 1M | 60 tok/s | Apache 2.0 |
Cached input is what a prompt token costs when the prefix was already on the card — the same text you send again on the next call. The API reports how many tokens landed there, and only those are billed at the cached rate. Decode speed is one stream on one GPU; concurrent streams share the card and move together.
Dedicated GPUs
Or rent the card and stop paying per token
Hourly instead of per-token, charged on elapsed time in whole minutes and charged only while the deployment is up. Spot is cheaper and can be reclaimed by the provider; on-demand cannot.
| Card | Memory | Bandwidth | On-demand / hr | Spot / hr |
|---|---|---|---|---|
| B300NVIDIA | 288GB HBM3e | 8 TB/s | $24.03 | $10.66 |
| B200NVIDIA | 180GB HBM3e | 7.7 TB/s | $16.20 | $7.08 |
| H200NVIDIA | 141GB HBM3e | 4.8 TB/s | $4.72 | $2.02 |
| H100NVIDIA | 80GB HBM3 | 3.35 TB/s | $3.37 | $1.48 |
| A100NVIDIA | 80GB HBM2e | 2 TB/s | $2.97 | $1.29 |
| RTX-PRO-6000NVIDIA | 96GB GDDR7 | 1.8 TB/s | $1.88 | $0.94 |
| RTX-5090NVIDIA | 32GB GDDR7 | 1.79 TB/s | $0.94 | $0.40 |
| RTX-4090NVIDIA | 24GB GDDR6X | 1.01 TB/s | $0.67 | $0.26 |
| CMP-170HXNVIDIA | 64GB HBM2e | 1.49 TB/s | $0.19 | $0.14 |
Committed capacity
Buying enough of it to name your own rate.
Reserved capacity
Capacity held for you rather than pooled, so a surge is a scheduling problem we already solved.
Volume pricing
Committed monthly spend moves the per-token rate off the published sheet above.
Priority support
One queue, and a committed account sits at the front of it rather than the back.
Custom models
Fine-tunes and quantisations we serve for you alone, on the same endpoint.
Questions
What people check before they send traffic
Nothing here is a policy we hope you don’t read. The billing code and the privacy policy say the same things.
Is it really uncensored?
There is no filtering stage in the path to remove. The gateway resolves a model, sends your prompt to it, and returns what came out — no output classifier, no content check on the way back. Uncensored fine-tunes are in the catalogue above, and you can fine-tune your own on a training job.
Do you log my prompts?
No. Every request records the model, the token counts and the cost — never the text of the prompt or the completion. That is also what makes the uncensored guarantee checkable: there is no stored text for anyone to read, sell, or train on. Debug logging is opt-in per API key, and debug logs are kept for 30 days.
What does a token actually cost?
Whatever the table says for that model. Fresh input, cached input and output are three separate rates, and a single response can carry all three. A non-streaming response also returns the cost of that exact request on an x-moton-cost-usd header, so you can price your own workload against our numbers instead of trusting an estimate.
What happens if my balance runs out?
Serverless calls are refused with a 402 before anything runs, so you never owe for a completion you could not pay for. A dedicated GPU is the opposite shape — the meter runs whether or not you are watching — so deployments stop automatically once the account is past its debt limit rather than letting the loss grow.
Can I run a model you don't list?
Yes. Point a deployment at it and you are billed by the GPU-hour rather than by the token; your own deployments answer on the same endpoint as the catalogue models. Anything that fits in the VRAM you rent is fair game, including fine-tunes you trained here.
What are the base model licenses?
Everything we serve on shared capacity today is Apache 2.0, and the licence is printed next to every model in the table above. Anything carrying a revenue or usage-triggered clause gets cut from the shared catalogue rather than passed on to you — if a model is on this page, its licence is the boring kind.
Do you charge per seat, or take a platform fee?
Neither. Flat plans are one monthly price for a plan, not one price per person, and serverless is billed on tokens at the published rate. Teammates on an account share the balance; nobody is metered for existing.
Start