One API. Every Model.
LrSite.h__desc__range
We publish what each model costs, where it runs and how it behaves before you write a line of code. No opaque reseller margin, no guessing which route you actually hit — you choose the model and see the price.
Context window, billing unit, p50 latency and the exact model ID are on the page. What you read is the route you get — we don’t silently swap providers mid-request.
Open the catalog →Your workload changes, so the model should too. Point a different model ID at the same request body and it just works — no client rewrite, no second integration.
Compare families →Every page ends in a request you can paste, not a demo you can watch. Copy the snippet, drop your key, and you have a live call.
Open the API reference →Video, image, text and audio all answer on the same host with the same key. Filter by what the model actually controls — resolution, duration, edit precision, context — then sort on price.
Clip generation from 480p to 4K, plus first/last-frame pinning, multi-reference input and native audio. Priced per second of output, so a 5-second spot costs five times the rate — not a flat guess.
Text-to-image and instruction-led editing, with reference images to hold a character or product steady across a series. Sort by resolution, edit precision and price per render.
Chat, reasoning, code and tool-calling agents with structured output. Compare on context window, output speed and the input/output split — the two numbers that decide your bill.
Full tracks, scored beds, voice-over, multi-speaker dialogue and sound design from a prompt — billed per track or per 1,000 characters of narration.
Routing internals, model comparisons and pricing notes from the team building the pipe.
Zhipu’s anonymous Ox Alpha topped OpenRouter for six days before its identity was revealed. We aggregate 4 independent reviews on pricing, benchmarks, multimodal reasoning and coding — then put it head to head with Claude Opus 4.8.
4 independent reviews on DeepSeek V4-Flash: pricing and cache economics, benchmark standings, long-horizon agent work and Vision-Exp multimodal — measured against V4-Pro and the overseas flagships.
5 independent reviews on xAI Grok 4.6: pricing, benchmarks, long-horizon agent runs and 3D/SVG/product-page generation — tested head to head against Claude Opus 5.
Three things we measure and publish, not promise.
Endpoints get probed every 900ms. One that starts to slow is pulled from the pool and your call is re-sent elsewhere — before it errors, not after.
Every request is itemised by model, key and route. Set a hard cap per key and you get warned while there is still credit left to move.
Tell the gateway what matters on this call — cheapest, fastest or most stable — and it picks the endpoint per request. You can also pin a route when you need determinism.
Provider list price sits next to the LeapRiver price on every model, before you call. You are billed per unit with no platform fee and no recurring charge.
Email and a password. $1 of credit lands immediately so you can call a real model straight away — no card details.
The key works for every model on the platform, and you can issue a separate one per environment with its own spend cap.
The endpoint speaks OpenAI, Anthropic and Google wire formats, so you change the base URL rather than rewriting your client.
From the first call you get per-request routing, automatic failover and the routed price rather than list.
The short answers, before you open a support ticket.