RouterStack sits between your app and 300+ AI models. One OpenAI-compatible endpoint, one key, one bill — with cost-aware routing and automatic failover when a provider goes down.
RouterStack speaks the OpenAI wire format. Point your existing OpenAI SDK at our base URL and every model in the catalog becomes a string you can swap at runtime.
from openai import OpenAI client = OpenAI( base_url="https://api.routerstack.ai/v1", api_key="rs-...", # one key for everything ) resp = client.chat.completions.create( model="anthropic/claude-opus", # or any of 300+ messages=[{"role": "user", "content": "Hello"}], extra_body={ "route": "cheapest", # or "fastest" "fallbacks": ["openai/gpt-5.2", "meta/llama-4-70b"], }, )
Set base_url to api.routerstack.ai/v1. Streaming, tool calling, JSON mode, vision — all pass through unchanged.
Model selection becomes a runtime decision, not an architecture decision. A/B a new model by changing one string.
If a provider is down or rate-limited, the request reroutes mid-flight to your next choice. Your users never see the outage.
Tokens, spend, latency, and error rates per model, per key, per app — in one place instead of nine provider consoles.
Frontier models from every major lab, plus open-weight models served on our own GPU fleet — RTX Pro 6000 Blackwell and H200 nodes — often at the lowest per-token price in the catalog.
| Model | Provider | Context | Input / Output per 1M |
|---|---|---|---|
| openai/gpt-5.2 | OpenAI | 400K | $1.25 / $10.00 |
| anthropic/claude-opus | Anthropic | 200K | $5.00 / $25.00 |
| google/gemini-3-pro | 1M | $1.25 / $10.00 | |
| meta/llama-4-70b | RouterStack GPU | 128K | $0.18 / $0.42 |
| deepseek/deepseek-v4 | DeepSeek | 160K | $0.26 / $1.10 |
| qwen/qwen3-235b | RouterStack GPU | 128K | $0.20 / $0.60 |
| mistral/mistral-large-3 | Mistral | 128K | $1.80 / $5.40 |
Every team that ships AI to production eventually learns the same lesson: single-provider is a single point of failure. RouterStack makes multi-provider the default, not a rewrite.
Provider 529s, rate limits, and timeouts trigger instant rerouting to your declared fallbacks. Zero code in your app, zero downtime for your users.
Route "cheapest" sends each request to the lowest-price provider that meets your latency SLO. Teams typically cut inference spend 30–60% without touching quality.
Route "fastest" picks the provider with the best live p50 from your region, measured continuously — not from a stale status page.
Restrict which prompts may reach which providers. Pin sensitive workloads to zero-retention endpoints or to our own GPU fleet only.
Per-request logs, token counts, spend, and latency histograms across every model — exportable, with spend alerts before the bill surprises you.
Open-weight models run on RouterStack's own GPU infrastructure in tier-3 data centers — not resold capacity. When we quote a price, we control the stack behind it.
No subscription, no platform fee at launch, no minimums. Buy credits, spend them on any model in the catalog.
We're onboarding in batches. Early access teams get founder pricing locked for 12 months and $50 in launch credits.