Priced by where it runs.
Our accelerators bill per token. Yours bill per EIM-hour, plus a rate per uncached token. Your datacenter bills per node. Hosted and Self-Hosted usage draws on prepaid credits: how credits work.
Rates are shown to signed-in accounts. Open an account to see them, or email contact@erebine.ai.
Three ways to run
Hosted
Our accelerators run your models. Pick a tier per endpoint; Free bills nothing.
Open an accountYour accelerators
The Self-Hosted tier: our router, metering and workspace in front of EIMs you run. You pay for the hours each EIM is connected and for the uncached tokens your EIMs serve; cached input tokens are free.
Enroll an EIMSovereign
The whole stack in your datacenter, licensed per node. EEMs included. Six-month minimum.
Sovereign termsHosted tiers
A tier sets an endpoint's rate, limits and scheduling priority. It never picks the hardware. Tiers without live capacity are not listed.
| Tier | For | Billed |
|---|---|---|
| Free | Evaluation limits at no charge. | No charge |
| Bulk | Lowest metered rate. Scheduled behind the interactive tiers; built for volume and background work. | Per token |
| Base | Entry metered rate for interactive traffic on the shared fleet. | Per token |
| Plus | Mid-ladder metered rate for interactive traffic on the shared fleet. | Per token |
| Max | Top metered rate for interactive traffic on the shared fleet. | Per token |
Endpoint, model-size and timeout limits per tier: Service tiers.
Compare with running it yourself
Two questions: when your own accelerators beat per-token billing, and what you would build to run without us.
- An OpenAI-compatible API: chat completions, Responses, embeddings, rerank, files, batches, images and audio, in the shapes the SDKs expect.
- A scoring router that weighs sixteen signals for a warm prompt and for any request your weights or a standing placement do not place, plus queueing and SLO tracking.
- An inference mesh of EIMs on NVIDIA, AMD, Intel, Apple silicon and CPUs, enrolled with join keys over a mutually authenticated CurveZMQ mesh.
- Metering and a credit ledger: per-token, cached-token and hourly billing, plans and budget alerts.
- Multimodal: image and video input, audio input, transcription, speech, realtime transcription and image generation.
- The EEM execution plane: tool calls on your hosts, with approvals before writes and destructive calls.
- Safeguards: AES-256-GCM at rest under per-project keys, project isolation, rate limits, scoped and expiring API keys, roles with two-factor sign-in, an audit log, zero data retention per workspace, and abuse, DMCA and NCII takedown intake.
Each line above is a project of its own. Together they are years of an experienced team's work, and an on-call rotation from the day it serves traffic.
Also billed
- Storage: one pool per project, the same rate on every tier. Storage billing
- Cached tokens: billed below the tier's token rate on hosted tiers, free on Self-Hosted. Prefix caching
- Plans: Each grants credits every cycle, and the larger ones bill tokens below the list rate. Plans
- Request lanes:
prioritybills above the token rate andflexbelow it. service_tier - Erepress: trims input tokens before they are billed, on every tier, at no charge. Erepress
- Rate limits: the same on every tier and plan. Rate limits
The whole stack, in your datacenter.
The build that runs erebine.ai, licensed per node for your network and your accelerators. Billed monthly, six-month minimum; term length, isolation guarantees and add-ons are set per contract. What the license includes.