OrcaRouter grades every prompt and routes it intelligently. Frontier-quality AI at up to 40% lower cost. Adaptive routing, load balancing, guardrails, agent firewall, observability, and governance — all through a single OpenAI-compatible endpoint.
No credit card · live in 60 secs
- client = OpenAI(api_key="sk-...")+ client = OpenAI(+ base_url="https://api.orcarouter.ai/v1",+ api_key="sk-orca-..."+ )# Everything else stays the same.response = client.chat.completions.create(model="orcarouter/auto", # router picks the best model per requestmessages=[{"role": "user", "content": "..."}])# → orcarouter/auto grades the prompt → frontier or open-source, zero token markup ✓
One line. We grade each prompt, route to frontier or OSS, and add $0.
Drop-in OpenAI-compatible, or connect agents over the OrcaRouter MCP server — keep your SDK, framework and editor.
Smart routing and automatic failover on every request.
Every prompt is embedded and routed by a model that keeps learning online from real traffic. On the public RouterArena leaderboard (Jun 2026) it leads on accuracy — ahead of GPT-5, Azure, Martian and NotDiamond — at 75.5%.
When a provider rate-limits or 5xxs, OrcaRouter retries the request against a healthy model across 200+ options before the response starts — so transient upstream outages don't surface to your users.
orcarouter/auto is a smart default, not a black box. Point each workspace at the objective you want — the cheapest model that clears your quality bar, the highest quality, or a balance of both — or let it learn the trade-off from your own traffic. You're never locked into one behavior.
When the default isn't enough, express routing as code — version-controlled, reviewable, and live in seconds. No redeploy, no client change.
version: 1rules:- id: hard_agent_taskwhen: task_class == "agent" && difficulty >= 0.6use:model: "claude-opus-4-7"reasoning_effort: high # spend where it matters- id: short_promptswhen: request.input_tokens < 500use: { delegate: cheapest }default:delegate: balanced # fall back to the chosen mode
See and prove every call — cost, model, latency, and why.
See exactly what every request cost, which model served it, how long it took, and why it failed — full structured logs you can filter, replay, and copy as a runnable cURL. A route is never a black box.
You pay each provider their exact price — we add $0 per token, ever. Every request shows the grade, chosen model, provider, latency, and price, so cost is glass-box, not an opaque blended rate.
Versioned prompts and caching — without a redeploy.
Version prompts behind named labels with A/B splits and one-click rollback. Move a label and every request picks it up instantly — no redeploy, no code change, no client update.
Repeated and cached prompt tokens bill at the provider's cache rate — often a fraction of the input price — across 5-minute and 1-hour ephemeral windows. Same answers, less spend, with cached_tokens on every receipt.
Guardrails, budgets, and an agent firewall that enforces.
PII Shield and content policies run before the upstream call is billed. A blocked request returns a clean 400 and is never charged — guardrails enforced inline, not logged after the fact.
Budgets and roles for people; a risk-scored firewall for agents. Every tool and MCP call is graded ALLOW, REVIEW, or BLOCK before it runs, and anomaly detection flags rate and cost spikes against learned hour-of-week baselines.
One URL change. Your existing SDK, model names, and streaming all work exactly as before.
Set base_url to api.orcarouter.ai/v1 and swap your API key. No other code changes needed.
Every call is routed to the best model, checked against your guardrails, and metered — graded in under 1ms, with failover, caching and full logs built in.
Traffic goes direct to each provider's first-party API at their published rate — we add $0 per token. One OpenAI-compatible endpoint for routing, observability and governance.
200+ models with live, side-by-side pricing — what you'd pay the provider directly. We add $0 on top.
| Model | Routed to | Input /M | Output /M | Context | Quality |
|---|---|---|---|---|---|
| kimi/kimi-k3NEW | Moonshot | $3.00 | $15.00 | 1M | 9.0 |
| openai/gpt-5.6-lunaNEW | OpenAI Direct | $1.00 | $6.00 | 1M | 7.0 |
| openai/gpt-5.6-terraNEW | OpenAI Direct | $2.50 | $15.00 | 1M | 8.0 |
| openai/gpt-5.6-solNEW | OpenAI Direct | $5.00 | $30.00 | 1M | 9.0 |
| grok/grok-4.5NEW | — | $2.00 | $6.00 | 500K | 9.0 |
| tencent/hy3NEW | — | $0.180 | $0.590 | 262K | 8.0 |
| anthropic/claude-sonnet-5NEW | Anthropic Direct | $2.00 | $10.00 | 1M | 9.0 |
| kling/kling-3-turbo | — | $0.112 /call | — | — | — |
| + 194 more models · Prices update every 60 seconds | |||||
We never take a cut of your token spend. Our revenue comes from optional team features.
What we're building and why — the five latest posts.
Short, factual answers — the same set search engines see in this page's structured data.
An AI router sits between your app and many models and picks the best model for each request. OrcaRouter grades every prompt in under 1 ms and routes it — frontier for hard reasoning, open-source for routine — across 200+ models at zero token markup.
An LLM router classifies each prompt and matches it to the model most likely to answer well at the lowest cost. OrcaRouter routes on contextual embeddings with online learning from live traffic — 75.5% accuracy on the public RouterArena leaderboard.
Yes. OrcaRouter is a production AI gateway: one OpenAI-compatible endpoint with adaptive routing, load balancing, automatic failover, guardrails, an agent firewall, prompt caching and per-request observability.
Like a CDN serves content from the best edge, an AI CDN serves inference from the best provider: healthy, fast, cheap capacity, cached repeated prompts, failover on outages. OrcaRouter plays this role across 200+ models.
Adaptive routing learns the quality/cost trade-off from your own traffic. Point each workspace at cheapest-that-clears-the-bar, highest quality, or balanced — or let orcarouter/auto keep tuning the choice per request.
They solve different problems — governance vs model choice — but you don't need two systems: OrcaRouter routes per request and enforces budgets, guardrails and observability on the same hop.
Prompt grading takes under 1 ms and total added latency stays under 50 ms — usually won back many times over by faster providers and cached prompt tokens.
No. You pay each provider's published rate; OrcaRouter adds $0 per token and monetizes optional Team and Enterprise features.
Yes — switch base_url to https://api.orcarouter.ai/v1 and your OpenAI, Anthropic or Google SDK code keeps working: Chat Completions, Responses, Embeddings, Images, Audio and streaming.
OrcaRouter retries against healthy fallback capacity for the same or an equivalent model before the response starts, so upstream outages don't surface to your users.
Swap one line. That's the migration.
Sign up with GitHub — no credit card required. You're live in under a minute.
OrcaRouter