OpenRouter

Core capabilities

Multi-provider routing is the base. When several providers serve the same model, OpenRouter load-balances on price unless told otherwise: it drops providers with significant outages in the last 30 seconds, picks among the cheapest candidates weighted by the inverse square of price, and keeps the rest as fallbacks. Developers can also sort by price, throughput or latency, pin or exclude providers, and set a maximum price.

Automatic model selection. openrouter/auto picks a model per request, and an August 10, 2026 update made it follow the past 7 days of real usage on the platform, which the post puts at more than 55 trillion tokens of weekly spend (as reported by the vendor). Fusion sends one prompt to a panel of models and has a judge model merge the answers; the vendor's tests on the DRACO deep-research benchmark report panels scoring above single models (as reported by the vendor).

Cheaper paths. The batch API launched on September 22, 2026: providers complete requests at a time of their choosing within 24 hours and usually charge 50% of their normal per-token price. The vendor says the median batch in its two-week beta finished in 7 minutes, across 230,000+ batches (as reported by the vendor).

Enterprise features keep arriving. In-region routing decrypts and forwards requests only inside a chosen region and returns an error rather than leave it; Security Center lists every key with its risk; the activity dashboard splits spend by agent, model and request; and private endpoints, launched October 14, put an enterprise's own fine-tuned models or dedicated deployments behind the same API, visible only to that organization.

Each of these features has its own rules. In-region routing uses separate base URLs: requests to us.openrouter.ai are decrypted inside the US and go only to US provider endpoints, while eu.openrouter.ai covers the EU and EEA, the UK and Switzerland; API keys, request bodies and model IDs stay the same. Security Center's overview sorts keys into four groups: no spend limit, never expires, unused or idle for 90+ days, and safe to remove. The vendor says its own audit found more than 1,000 active keys across 85 employees (as reported by the vendor). Private endpoints are an Enterprise add-on, stay out of public model lists, rankings and benchmarks, and can fall back to public providers if the endpoint fails or hits its rate limit. The activity dashboard opens with five metrics: total spend, requests, token volume, cache hit rate and blended cost per million tokens, and the same data is available through the Analytics API.

Public data. The rankings order models and apps by real token usage and are one of the data sources commonly cited on model market share. The main leaderboard ranks by tokens processed and the market-share chart counts requests; the page states that neither measures users, spend or task quality. The rankings data is licensed CC BY 4.0.