The largest single-API model catalog on Earth — 26,116 models across all 12 modalities — chat, image, video, 3D, music, speech, embeddings and more — all behind one OpenAI-compatible endpoint. Zero logs. Zero rate limits. Prompts never stored. Pure PAYG per-model rates.
We don't rent you a GPU. We intelligently route requests across a shared fleet of GPUs, loading and unloading models in real-time. Every millisecond of compute is utilised — no idle hardware, no wasted spend. You pay for inference, not for sitting empty.
Thousands of users share the same model instances simultaneously. A single 70B model serving 500 concurrent requests costs a fraction per-user compared to dedicated hardware. This is how we deliver frontier models at per-model PAYG rates instead of £1.59/M.
Per-model, per-unit billing — per M tokens, per image, per video, per second of audio. Top up any amount from £10, pay only for what you use at published per-model rates from £2e-07/M. Cache reads 65% off input. No subscriptions, no tiers, no expiry — every request metered to the exact unit.
Prompts and completions are never stored, logged, or transmitted to third parties. Ephemeral by design. No telemetry. No usage analytics. GDPR compliant by default.
26,116 models on pure PAYG — frontier chat flagships, image, video, 3D, music, speech, embeddings and more, all 12 modalities behind one OpenAI-compatible API. One key, one bill.
Popular models stay hot in GPU memory. Rare models load on-demand in seconds, then free their resources. You get instant access to 26,116 models without paying for 26,116 GPUs. Only the inference you use, nothing more.
Point your existing OpenAI SDK at api.openrout.ing and your plain chat completions endpoint gains a full intelligence runtime: context enrichment, tiered memory, knowledge graphs and workflow orchestration — all behind one key, billed as normal inference.
Every request can pass through an enrichment pipeline that injects relevant memories, knowledge and facts before inference — your prompts arrive pre-loaded with context the model actually needs. Watch it work via the X-Enrichment-Ms and X-Enriched response headers. Enable with "enrich": true.
Upload documents or paste raw text — the system auto-extracts entities and relations, builds a queryable knowledge graph, and stores 347K files for retrieval. Structured fact queries via dedicated endpoints, not just unstructured RAG. Build: POST /knowledge/build. Query: GET /knowledge/graph/query.
Every knob you'd build yourself, already built: model aliases with coalition fallback (41 model mappings, 3 retries), per-client concurrency slots, per-tier rate limits, structured output & JSON mode, seed reproducibility, logprobs, streaming SSE, batch inference, and response headers (X-Tier, X-Model-Used, X-Coalition, X-Queue-Wait, X-Request-ID) that expose exactly what the router did with your request.
One API key. Two endpoints. raw.openrout.ing/v1 gives you direct OpenAI-compatible completions — pure model access, no orchestration, minimum latency. api.openrout.ing/v1 wraps every request in a full intelligence runtime: context enrichment, persistent memory, workflow orchestration, knowledge retrieval, guardrails, and multi-step reasoning — all included free.
raw.openrout.ing/v1
api.openrout.ing/v1
Every model, every modality — chat, image, video, 3D, music, TTS, embeddings — billed per request at per-model PAYG rates from £2e-07/M. No subscriptions, no tiers, no lock-in. Credits never expire.
Sign up, grab a key, top up PAYG credits. No sales calls, no KYC for standard plans, key works instantly.
Change one base URL in your existing OpenAI SDK. Raw inference on raw.openrout.ing, intelligent inference on api.openrout.ing — same code otherwise.
26,116 models by ID — or use aliases and auto-routing to always hit the best healthy model for the job. Coalition fallback included.
Enable memory, auto-summarization, knowledge graphs and workflows through request parameters. No rewrite — enrichment activates on the same endpoint, and conversation summaries are generated automatically so long threads stay cheap.
Zero 429s, queue-indefinite rate limiting, per-tier concurrency. Your traffic grows; your integration never changes.