OrcaRouter is a production-oriented OpenAI-compatible AI gateway (AI Gateway for Adaptive Routing) that integrates adaptive routing, load balancing, guardrails, an agent firewall, observability, and governance into a single endpoint, covering more than 200 models. Its core philosophy is "one gateway, all models, your data always stays yours."
Purpose and Positioning OrcaRouter sits between applications and multiple models, acting as an AI gateway / AI router / AI CDN: it scores the prompt of each request (claimed to be under 1 millisecond) and then routes it to the model best suited to answer that request—hard reasoning goes to frontier models, routine tasks go to open-source models. It simultaneously enforces budgets, guardrails, and observability on the same hop.
Main Capabilities
- Routing and load balancing: adaptive routing based on contextual embeddings and online learning, officially stated to achieve 75.5% accuracy on the public RouterArena leaderboard and to beat GPT-5 and Azure. It provides automatic failover, prompt caching, and per-request observability. Each workspace can be set to "cheapest that meets the bar," "highest quality," or "balanced," and orcarouter/auto can be used for per-request automatic tuning.
- Compatibility: one OpenAI-compatible endpoint supporting Chat Completions, Responses, Embeddings, Images, Audio, and streaming output; OpenAI, Anthropic, or Google SDK code can continue to be used, requiring only a switch of base_url.
- Privacy architecture: Data Cloaking replaces names, email addresses, and card numbers with stand-ins before the call and restores them in the response; ZDR zero retention, with prompts not logged by default unless logging is enabled, and a zero-retention agreement can be signed to lock it in; PQC post-quantum encryption, where logs that users choose to retain are sealed with hybrid ML-KEM, with the encryption open-sourced (SCUTTLE) and enabled by default; data residency can be declared as US, EU, UK, Asia-Pacific, or China, and compliance reports are marked accordingly; TEE trusted execution environment, with attested enclaves and a proof provided for each answer.
- Guardrails: intercept PII, keys, jailbreaks, and prompt injection before the model sees them, blocking them before billing occurs; pattern rules are free, and AI judge rules are billed only for the check itself.
- Agent firewall: scores every tool and MCP call, allowing, escalating to human review, or blocking before execution, with defaultRecognized as being in an observation state.
- Compliance packs: mapped to 32 frameworks, one policy per regulation, allowing observation on real traffic before enabling enforcement (paid plans).
- Human approval queue, tagging threats to OWASP LLM Top 10 and MITRE ATLAS, threat tagging based on OWASP LLM Top 10, public archive of agent incidents.
- Prompt version management and cache reuse, with no code changes required.
How to use Officially described as "live in 60 seconds": set the SDK's base_url to api.orcarouter.ai/v1 and replace the API key, with no other code changes required; routing, guardrails, and observability are handled by OrcaRouter.
Billing and pricing OrcaRouter claims zero markup on pay-as-you-go and subscriptions, with tokens billed at each provider's published price and no markup of its own; revenue comes from optional team features. Three payment methods: pay-as-you-go (top up anytime, at provider prices, with a receipt for each call), credit subscription (top up the wallet each cycle, full amount, any model, larger plans include bonus credits), and bring your own key (use your own provider keys, rate limits, and credits). Plans are aimed at team features and do not change token prices.
Use cases Suitable for production AI applications and teams that need unified access to multiple models, intelligent per-request model selection, cost control, and governance, security, and observability in the same layer.
Limitations and notes According to the official description, data region settings currently only label compliance reports and do not yet fix where data is stored or where inference runs; inference is executed by the provider corresponding to the selected model within its region. Regarding privacy, the official statement says prompts are not used for training, and unless logging is enabled for the workspace, prompt and response bodies are not recorded, but request metadata (time, model, tokens, cost, IP) is retained for billing and security. Prompt scoring takes less than 1 millisecond, and total added latency is less than 50 milliseconds. Officially, migration only requires changing one line.
The above information all comes from the official page; specific feature availability, regional support, and team feature details may vary by plan, and a more detailed version and region list is not currently provided.
