One LLM API gateway for every model

An LLM API gateway is a single endpoint you route every model call through, instead of wiring each model provider into your app one by one. Router One is an OpenAI-compatible gateway for 30+ models — GPT, Claude, Gemini, and Grok — behind one base URL and one key. Direct model calls are a black box; routing through a gateway gives you a ledger (what every call cost), a trace (which model and route served it), and control (per-key budgets and rate limits). It is reachable globally and from Mainland China.

Get your API key

What the gateway does

One OpenAI-compatible endpoint

Point your existing OpenAI SDK at https://api.router.one/v1 and call 30+ models — GPT, Claude, Gemini, Grok — through the same Chat Completions interface. No per-model SDK, no rewrite.

Smart routing

Production traffic uses the model_name strategy by default, so a named model stays fixed. Send model="auto" to use server-side candidates and a global retry budget. Optional adaptive signals include latency, cost, and reliability, with time-decayed EWMA scoring.

Automatic fallback

When a named-model request hits a retryable upstream error, the gateway may try the next healthy provider serving that same model. It does not silently switch to a different model.

Per-request traces

Every request is traced in a real-time dashboard: final model and provider, token counts, cost, latency, and status. Internal failed attempts are not presented individually.

Keys with budgets + limits

Create multiple API keys per account, each with its own maxSpend cap, rate limit (rateLimit), and per-minute token ceiling (tokenLimitTpm). Give each app or environment its own scoped key.

Prepaid pay-as-you-go wallet

Top up a prepaid wallet and pay per token at the posted model-rate token line, with checkout-visible FX/channel fees kept separate. Fund it with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains. Spend draws down as you go.

Native Claude Code & Codex

Claude Code talks to the Anthropic-compatible endpoint at https://api.router.one; OpenAI Codex CLI uses the OpenAI-compatible base URL. Claude Code switches with one environment variable; Codex CLI takes a three-line provider block in ~/.codex/config.toml (base_url, wire_api = "responses", env_key) plus the key it names.

Global + China reachable

The same gateway is reachable from global networks and from Mainland China, so one integration serves teams on both sides without a separate setup.

Ledger, trace, and control

The gateway sits between your app and the models so it can do three things a direct call can't. The ledger records what every request cost at posted rates, drawn from your prepaid wallet. The trace records which model, provider, and route served each request, with tokens and latency. The control layer enforces per-key budgets and rate limits before a request reaches a provider, so a runaway loop or a leaked key can't drain the wallet.

Switch with one base URL change

Router One speaks the OpenAI Chat Completions API, so most integrations move over by changing the base URL and key. Set the two environment variables and your existing code keeps working.

terminal

# Route every model call through the gateway
export OPENAI_BASE_URL=https://api.router.one/v1
export OPENAI_API_KEY=sk-your-router-one-key

FAQ

What is an LLM API gateway?

An LLM API gateway is one endpoint you route every model call through instead of integrating each provider separately. It gives you a unified interface for 30+ models, plus routing, fallback, cost and latency traces, and per-key budget controls in one place — a ledger, a trace, and a control layer over your model usage.

Is Router One OpenAI-compatible?

Router One is OpenAI-compatible: it implements the OpenAI Chat Completions API at https://api.router.one/v1, so any library or tool that works with OpenAI works by changing the base URL. Claude Code uses the Anthropic-compatible endpoint at https://api.router.one.

Which models can I call?

30+ models across the major families — GPT, Claude, Gemini, and Grok — all through the same unified endpoint. Browse the current catalog and per-model rates on the models page.

How does routing decide which model to use?

The default production strategy is model_name, which honors the model you name. Send model="auto" to let the server select from its candidate set under a global retry budget. Optional adaptive signals include latency, cost, and reliability, with time-decayed EWMA scoring. The dashboard shows the final model and provider.

What happens when a provider goes down?

For retryable upstream errors, a named-model request may be retried on the next healthy provider serving that same model. The dashboard shows the final provider and request metrics, not every internal failed attempt.

How does billing work?

You top up a prepaid wallet and pay per token at the posted model-rate token line, with checkout-visible FX/channel fees kept separate. Top up with a card or Alipay through one hosted checkout, or with USDT/USDC on six chains. Each API key can carry its own spend cap and rate limit so usage stays within budget.

Choose the next step for your gateway evaluation

This page explains the gateway layer. Use the paths below to test an existing client, understand routing decisions, follow request costs, inspect the route that served a request, compare gateways against relay platforms, or open a client setup guide.

MigrationMove an existing OpenAI clientUse the exact base URL and key swap, then confirm the SDK-compatible request contract before changing production traffic.Request contractCheck structured outputs through the gatewayWhat the gateway validates in response_format before a model is called, what it forwards as sent, and why enforcement stays with the model.RoutingUnderstand model and provider selectionSee how named-model requests, auto selection, candidate sets, retry budgets, and adaptive signals fit together.Cost controlTrack request and API-key spendFollow token usage and billed cost from an individual request into the API-key and account-level views.VerificationInspect the final route, latency, and costUse request traces to verify the final model and provider, token counts, latency, status, and billed cost.EvaluationCompare gateways against relay platformsFive concrete differences between a managed gateway and a stitched-together API relay: stability, traceability, and billing.ReliabilitySee what happens when a route failsHow retryable provider errors (5xx, timeouts) can be retried on another healthy route serving the same requested model — conditional, not a zero-downtime guarantee.Client setupConnect your client or SDKOpen the per-client integration guides to point an editor, CLI, or SDK at the gateway endpoint.FrameworkConfigure LlamaIndexPoint LlamaIndex's OpenAI-compatible LLM class at the gateway base URL and switch models by string.Model catalogCopy a current model ID and its rateThe live catalog lists every model ID the gateway serves, its per-1M-token rate, context window, capabilities, and the endpoints it answers on.

Route every model through one gateway.

Get your API key