BUILD ON RIZON

Your agent's next step.

Pair open-model inference with a persistent Linux workspace. Use either product on its own, or put them together for agents that work over hours and days.

Inference quickstart

Rizon supports OpenAI-compatible Responses and Chat Completions APIs, plus an Anthropic-compatible Messages endpoint. Use your existing SDK with a Rizon API key and endpoint.

Create an API key in your Rizon console. Keep it on your server.
Set the RIZON_API_KEY environment variable.
Choose an available model ID from the console and replace YOUR_MODEL_ID below.
Send a request. Set the service tier explicitly when your workload needs a specific cost and latency tradeoff.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.rizon.live/v1",
    api_key=os.environ["RIZON_API_KEY"],
)

response = client.responses.create(
    model="YOUR_MODEL_ID",
    input="Research the next generation of battery storage.",
    service_tier="flex",
)

print(response.output_text)

Automatic prompt caching keeps eligible input available for 20 minutes. Cached input is billed at a lower rate. Repeated prefixes work best: keep changing values, such as timestamps, out of the start of your prompt.

Retry with care. A 503 can arrive after a request was accepted. Use an idempotency key to avoid running and billing the same work twice.

Choose a service tier

Set service_tier on a request. The response reports the tier that actually served it, and billing follows that returned tier.

Inference service tiers
Tier	Use it for	Keep in mind
`fast`	Latency-sensitive work	Higher cost. Requests with custom tools are rejected.
`default`	Everyday agent work	Requests with custom tools use this tier unless a tier is set.
`flex`	Background work and cost-sensitive jobs	Lower cost, with more waiting. Required for batch and LoRA adapters.
`auto`	The project's configured tier	The default request setting. Projects start on default, but an admin can change this.

Batch requests run asynchronously on flex. Match results by request ID and use webhooks to handle completion. Customer LoRA adapters also run on flex, on shared base models.

A persistent workspace

A RizonBox is a managed Linux computer for your agent. Run commands, read and write files, schedule work, and publish ports. Apps group boxes; a box can be private to its creator or shared with the organization.

Lifecycle and saved state

RizonBox lifecycle
Action	What to expect
Pause	Stops compute charges. Does not wake on traffic to a published port.
Sleep	Stops compute charges. Can wake on traffic to a published port or a scheduled wake.
Wake or resume	Can be cold: the disk remains, but memory and running processes may be lost.
Reboot or upgrade	Clears memory and keeps disk and volumes. On a paused or sleeping box, the action waits for the next wake and returns `applied: false`.
Checkpoint and fork	Save state and start another box from a checkpoint. Check the checkpoint's `expires_at` before relying on it.
Terminate	Final. Data that is not on a volume or in a checkpoint is lost.

Check the result, not only the HTTP status. A resume can return 200 with resume_state: terminal_unavailable. Check for running or already_running before sending more work.

A pending upgrade makes the next wake cold. After a runtime deprecation deadline, a box can restart to upgrade. Make agent progress durable and check runtime notices before long runs.

Storage, networking, and secrets

Volumes outlive individual boxes and are billed for every hour they exist. Use them for data that must survive termination. Connect your workloads with published ports, custom domains, tunnels, and SSH certificates.

Egress policies control which hosts a box can reach. Secrets are injected into requests for allowed hosts. When an external service cannot be reached, check the egress policy before retrying.

Billing that follows the work

Inference is billed by token at the rate of the tier that served the request. Cached input has a lower rate. Review the response's service tier and your key's token usage when reconciling a bill.

RizonBoxes are billed only while running, for used vCPU, memory, and disk, in one-minute increments. Volumes are billed for each hour they exist, including while boxes are paused or sleeping. Review minute-level spend and resource usage for each box in the console.

Monthly plan terms · USD
Plan	Price and credits	Capacity
Free	$0 subscription. $5 monthly credit with a payment method. Prepaid usage; creation fees apply.	4 seats, 100 concurrent boxes.
Pro	$250 per month with $100 monthly credit. Prepaid usage; creation fees waived.	Unlimited seats, 5,000 concurrent boxes, higher rate limits.
Enterprise	Custom volume pricing, invoiced monthly in arrears. Creation fees waived.	Unlimited seats and custom limits. Uptime and latency SLAs; dedicated Slack support.

Free supports email and community support; Pro includes email support. US-only inference is available on Pro at a 20% surcharge and Enterprise at a 10% surcharge. HIPAA support requires Enterprise and a signed BAA. Confirm the scope of your agreement before sending regulated data.

Compare plans →

Data and access

Support sees operational metadata: box states, resource metrics, request and token counts, and error codes. It cannot see prompts, completions, files, or command output. Accounts with zero data retention have no request logs.

If content is needed to investigate an issue, you can share it with support or agree to a logged owner override. An override requires a reason. Organization admins can also use an owner override to act on a private box.

Keep API keys and other credentials out of prompts and source control. Use host-scoped secrets and egress policies for services your agent needs to access.