LLM inference
built for production scale.
Run open-weight models on dedicated capacity tailored to your workload, or get started instantly with our on-demand Model API.
Two ways to run
LLM inference.
Start on shared capacity in minutes, or reserve capacity sized to your workload. Both run on the same infrastructure, so moving from one to the other is not a re-integration.
- On-demand
On-demand Model API
“I want to get started quickly, evaluate models and serve variable workloads.”
- Every major open-weight and proprietary model
- Usage-based pricing with prepaid credits
- Prompt caching, tool calling and structured outputs
- No subscription. Pay only for what you use
From $0.032 per 1M input tokens01 - Dedicated
Dedicated LLM inference
“I have a production workload and need capacity, performance and economics tailored to my scale and requirements.”
- Guaranteed, high-throughput capacity with increased rate limits
- Volume-based pricing with custom SLAs
- Custom models and workload-specific optimization
- Region-locked data processing and Zero Data Retention
Sized and priced for your workload02
Dedicated capacity,
sized to your workload.
When traffic is steady, latency is a requirement or data has to stay in one region, shared capacity stops being the right shape. Dedicated inference reserves infrastructure for your models and tunes it to the way you use them.
- Capacity
- Guaranteed, high-throughput capacityInfrastructure reserved for your traffic and built for real-time workloads, with rate limits raised to match what you reserve.
- Performance
- Inference tuned to your workloadWorkload-specific optimization raises throughput and brings down the cost of every token you serve.
- Economics
- Volume-based pricing, custom SLAsPricing follows your volume instead of a public rate card, with service levels agreed for your deployment.
- Models
- Open-weight models, or your ownRun the open-weight models you already rely on, or bring custom and fine-tuned models to the same capacity.
- Residency
- Region-locked data processingKeep inference inside one region when your data has to stay there, in the US or in the EU.
- Privacy
- Zero Data RetentionPrompts and outputs are processed in memory and not retained. Nothing you send is used to train or improve models.How Zero Data Retention works
LLM API pricing
per million tokens.
50 language models from 9 labs answer on one endpoint, each priced per token with no subscription. 11 of them are open-weight models running on Runware hardware.
| Model | Input | Output |
|---|---|---|
anthropic:[email protected] | $4.00 | $20.00 |
anthropic:claude@opus-5 | $5.00 | $25.00 |
| $0.75 | $3.75 | |
anthropic:claude@fable-5 | $10.00 | $50.00 |
moonshotai:kimi@k3 | $3.00 | $15.00 |
anthropic:[email protected] | $2.00 | $10.00 |
| $1.20 | $4.00 | |
| $0.15 | $0.50 |
Runware-hosted models are open-weight models running on our own hardware.
ZDR marks the models covered by Zero Data Retention.
Prices marked “from” are the lowest tier. Longer contexts or audio input cost more, and the full pricing lists every tier.
Keep your SDK.
Change two values.
The API speaks the OpenAI Chat Completions format and, for Claude models, the Anthropic Messages format. Change the base URL and the API key, then call any model in the catalog.
- LLMs on one endpoint
- 50
- Lowest price per 1M input tokens
- $0.032
- Largest context window, in tokens
- 1.1M
from openai import OpenAI
client = OpenAI(
api_key="your_runware_api_key",
base_url="https://api.runware.ai/v1",
)
response = client.chat.completions.create(
model="moonshotai:kimi@k3",
messages=[{"role": "user", "content": "Summarize this incident report in three bullets."}],
)
print(response.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
apiKey: "your_runware_api_key",
baseURL: "https://api.runware.ai/v1",
});
const response = await client.chat.completions.create({
model: "moonshotai:kimi@k3",
messages: [{ role: "user", content: "Summarize this incident report in three bullets." }],
});
console.log(response.choices[0].message.content);from anthropic import Anthropic
client = Anthropic(
auth_token="your_runware_api_key",
base_url="https://api.runware.ai",
)
message = client.messages.create(
model="anthropic:[email protected]",
max_tokens=1024,
messages=[{"role": "user", "content": "Summarize this incident report in three bullets."}],
)
print(message.content[0].text)curl https://api.runware.ai/v1/chat/completions \
-H "Authorization: Bearer $RUNWARE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "moonshotai:kimi@k3",
"messages": [{"role": "user", "content": "Summarize this incident report in three bullets."}]
}'- Streaming
- Tool calling
- Structured outputs
- Prompt caching
FAQs
The on-demand Model API runs your requests on shared capacity and bills per token, so you can start in minutes and follow variable traffic. Dedicated LLM inference reserves infrastructure for your workload, with guaranteed throughput, increased rate limits, volume-based pricing and custom SLAs.
When the workload is in production and you can size it: steady or high-volume traffic, latency targets you have to meet, a custom or fine-tuned model, or data that has to stay in one region. Evaluate models on demand first, then talk to us once you know the shape of your traffic.
On demand, you pay per million input and output tokens from prepaid credits, with no subscription. Models with prompt caching bill cached input tokens at a lower rate. Every rate is on the pricing page. Dedicated inference is priced by volume for your workload.
50 language models from 9 labs, from open-weight models hosted on Runware hardware to proprietary models from the leading labs. The model directory lists every one with its context window and price.
Yes. The API accepts the OpenAI Chat Completions format, so the OpenAI SDKs work after you change the base URL and the API key. Claude models also answer in the Anthropic Messages format. The Compatible APIs reference covers both.
Yes, on dedicated capacity. Dedicated LLM inference serves open-weight models and your own custom models, with the serving stack tuned to your workload.
Runware never uses data sent through the API to train or improve models. Zero Data Retention is available for enterprise organizations on covered models, and dedicated deployments can lock processing to the US or the EU. See how Zero Data Retention works.
The on-demand API runs on shared capacity, with a per-organization rate limit set far above typical traffic. Dedicated capacity comes with increased rate limits that match what you reserve.
Start on demand. Scale on dedicated.
Call your first model in minutes, or tell us about your workload and we will size dedicated capacity for it.