A fixed price per model. All of it to the machine that did the work.
Planned reference prices for the alpha, per million tokens unless stated. Low because no data centre is rented and Rutli keeps none of the compute price. Set with the founding nodes and adjusted once the network is live.
PLANNED not an offer yet
Language and vision models
| Model | Kind | Input | Output | Runs |
|---|---|---|---|---|
| gpt-oss 20B | chat | $0.03 | $0.10 | Node 0001, today |
| gpt-oss 120B | chat | $0.06 | $0.25 | Node 0002, today |
| Qwen3-VL 30B | vision and chat | $0.05 | $0.18 | Node 0002, today |
| Llama 3.3 70B | chat | $0.10 | $0.30 | planned |
| nomic-embed-text | embeddings | $0.004 | n/a | both nodes, today |
Images and speech
| Model | Kind | Unit | Price | Runs |
|---|---|---|---|---|
| FLUX 2 image | image | per image, 1024 px | $0.008 | Node 0001, today |
| Whisper speech to text | speech | per minute of audio | $0.002 | Node 0001, today |
| Kokoro text to speech | speech | per million characters | $3.00 | Node 0001, today |
Per million tokens unless stated. "Runs" says where the model runs today on the founding nodes; "planned" means it does not run yet. Prices change only with a published date.
The big ones.
The models that matter most are the ones no single desk can hold: a trillion-parameter mixture of experts needs more memory than any one machine has. That is exactly what a network of owned machines is for. Split across many nodes, served at a fixed published price, none of it kept by Rutli.
PLANNED served across several machines once the network plane runs; the prices are the plan
| Model | Size | Rutli, input | Rutli, output | Elsewhere |
|---|---|---|---|---|
| Kimi K2.6 | 1T mixture of experts, open weights | $0.30 | $1.20 | Moonshot API: $0.95 in, $4.00 out (BenchLM and TokenMix, September 2026) |
| DeepSeek V3.2 | 671B mixture of experts, open weights | $0.12 | $0.20 | DeepSeek API: $0.21 in, $0.32 out (BenchLM, September 2026) |
| Qwen3 235B A22B | 235B mixture of experts, open weights | $0.15 | $0.35 | Alibaba Cloud: $0.46 in, $0.90 out (pricepertoken.com, 2026) |
| GLM-5 | open weights | $0.20 | $0.60 | Z.ai API: $0.60 in, $1.92 out (pricepertoken.com, 2026) |
| MiniMax M2.7 | open weights | $0.10 | $0.35 | MiniMax API: $0.24 in, $0.96 out (pricepertoken.com, 2026) |
Open-weight models only; the network runs what its owners can run. Elsewhere's figures are the providers' own or as reported on the date shown.
The same models elsewhere
Rutli's planned price beside each provider's own, per million tokens, on the same scale. They rent data centres and keep a margin. Rutli rents nothing and keeps nothing of the compute price.
- gpt-oss 20B 3.0× lower on output, 2.5× on inputGroq pricing as reported by CloudZero and AI Pricing Guru, 2026
- gpt-oss 20B 1.4× lower on outputArtificial Analysis provider comparison, 2026
- gpt-oss 120B 2.4× lower on output, 2.5× on inputtogether.ai/pricing, read 2026-09-05
- Llama 3.3 70B 3.5× lower on output, 10× on inputtogether.ai/pricing, read 2026-09-05
- Kimi K2.6 3.3× lower on output, 3.2× on inputBenchLM and TokenMix, September 2026
- DeepSeek V3.2 1.6× lower on output, 1.8× on inputBenchLM, September 2026
- Qwen3 235B A22B 2.6× lower on output, 3.0× on inputpricepertoken.com, 2026
- GLM-5 3.2× lower on output, 3.0× on inputpricepertoken.com, 2026
- MiniMax M2.7 2.7× lower on output, 2.4× on inputpricepertoken.com, 2026
Figures are the providers' own, as published or as reported on the date shown. If one changes, this page changes. Rutli's are planned reference prices for the alpha.
Questions
How is the price set?
Rutli publishes a fixed price per unit of work for each model: per million tokens for language models, per image, per minute of audio. There is no bidding and no surge. The prices on this page are planned reference prices for the alpha, set with the founding nodes and adjusted once the network is live; every change is published with its date.
Why is it this low?
No data centre rent. The work runs on machines people already own and are not using, and Rutli keeps none of the compute price, so there is no platform margin stacked on top. Rutli lives on the services around the network and says so.
Where does the money go?
A customer pays the published price. The node owner whose machine did the work receives 100% of it. Payment and payout costs are pass-through, at cost, shown separately.
How do I pay?
A balance in ordinary money, topped up with a card. Every line on it says what ran, where, for how long, and what it cost. No token to buy, nothing to hold.
Which models are available?
The ones running on the founding nodes today are marked as such in the table. Others are planned. A model appears here only when it runs.
Is my data logged?
A workload runs in a sandbox on the node, and the node keeps no copy of prompts or outputs. The control plane sees scheduling and metering, never content. Where a customer needs a jurisdiction, the network routes by it.
What about reliability and speed?
A node serves its owner first: a task of theirs preempts network work at once, and the job is resumed elsewhere. Latency and throughput depend on the machines available, and the status page publishes what is measured, never estimates.
When can I use it?
When the founding nodes run the first workloads for outside customers, in the alpha. Leave your address on the founding form and say you build software: the first mail is the workload contract for review.