Get Started
Concepts
The handful of ideas behind the platform: organizations, API keys, prepaid credits, and one OpenAI-compatible endpoint.
Control plane and data plane
The platform splits into two surfaces:
- The control plane, this dashboard and its management API: your organization, members, API keys, credits, billing, and usage analytics.
- The data plane, the inference endpoint your applications call. It is OpenAI-compatible, powered by GPU capacity across the Pearl network:
https://inference.pearlresearch.ai/v1Your SDK and API traffic goes straight to the data plane with a Pearl API key; nothing you ship routes through the dashboard.
Organizations and roles
An organization is the top-level resource: API keys, credits, usage, and billing all belong to it, and every member's usage draws from the same balance. You create one when you sign up, and invite teammates from the Organization page. Each member has a role that scopes what they can do: org-admin has full control, billing-admin covers credits and payments, developer covers day-to-day API work, and viewer is read-only.
API keys
API keys authenticate your requests to the data plane. They are organization-scoped, prefixed sk-prl-infapi-, and shown in full exactly once at creation, we store only a digest, so a lost key means minting a new one. Manage them on the API Keys page:
- Revoke disables a key. It stays listed as revoked so you can see what was shut off, and stops being accepted within a couple of minutes.
- Delete removes the key from your dashboard entirely. Usage it already generated stays attributed in your analytics.
Credits and billing
The platform runs on prepaid USD credits. Inference is metered per token, input, cached input, and output tokens are each priced per million, and drawn down from your organization's balance as you go. There are no subscriptions or minimums.
- New organizations receive a $5 welcome credit to start experimenting.
- Top up from the Billing page ($5–$4,999 per top-up) via Stripe-hosted checkout, card details never touch our servers.
- Optional auto-recharge tops the balance up automatically when it dips below a threshold you set.
- When credits run out, requests are declined until you top up. Enterprise organizations can instead be invoiced monthly, contact us to switch.
Models and the network
The catalog is a curated set of production open-weight models, see Choosing a model. Models are served on the Pearl network's distributed GPU capacity rather than a single provider's cluster, behind one OpenAI-compatible endpoint. The platform currently serves chat models; the live list (with per-model context lengths and supported parameters) is always at GET https://inference.pearlresearch.ai/v1/models.
OpenAI compatibility
Any OpenAI SDK works by changing two constructor arguments, the base URL and the API key. No other code changes are required.
| Endpoint | Status |
|---|---|
POST /chat/completions, including streaming, tools, JSON mode, and image input | Supported |
POST /completions, including streaming | Supported |
GET /models | Supported |
Request parameters
Models accept the standard sampling parameters, temperature, top_p, frequency_penalty, presence_penalty, stop, seed, and max_tokens, plus top_k and min_p extensions (pass those via extra_body in the OpenAI SDKs). Support varies slightly per model; the supported_sampling_parameters field on GET /models is authoritative.
Model IDs are namespaced (for example deepseek-ai/DeepSeek-V4-Pro), not flat names like gpt-4o, remember to update the model parameter when migrating.
Playground
The playground is an in-browser chat over the same catalog. Playground traffic is billed to your organization like any API call, so what you see there, behavior, latency, cost, is what your application gets.
Glossary
| Term | Definition |
|---|---|
| Token | The unit models read and write, roughly ¾ of an English word. Usage and pricing are metered per token. |
| Context window | The maximum number of tokens (prompt + response) a model can consider in one request. |
| Streaming | Receiving the response token by token as it is generated, instead of waiting for the full reply. |
| Function calling (tools) | Letting the model request that your code run a function, then continue with its result. |
| Structured outputs (JSON mode) | Constraining the response to valid JSON so it can be parsed by machines. |
| Reasoning model | A model that thinks before answering and returns that thinking in a separate reasoning field. |
| Cached input tokens | Prompt-prefix tokens the network has recently processed for you, metered separately at a discounted rate. |
| Quantization | The numeric precision a model is served at (fp8, int8), lower precision trades a little quality for speed and cost. |