OrcaRouter is an AI gateway from Continuum AI, a US company. It brings together more than 200 models from providers such as OpenAI, Anthropic, Google and DeepSeek so that you can call all of them through a single OpenAI-compatible API. What makes it distinctive is that it scores the difficulty of each incoming prompt on the spot and automatically routes easy requests to cheaper models and hard ones to high-end models. On top of that, it does not add a fee to token usage — it passes through each provider’s published rates as they are. Released in May 2026, it is a relatively new service, and it also supports connections from development tools such as Claude Code, Codex and Cursor.
Key Features
- Adaptive routing: Every request is scored automatically for prompt difficulty, and the gateway decides dynamically whether to send it to a frontier model or an open-source one. Your application stays the same while cost and quality are balanced per request
- Zero-markup billing: Token charges pass straight through at the providers’ published rates, with no fee layered on top. Revenue comes from subscriptions to the team-oriented features instead
- Automatic failover: If a provider goes down, the gateway retries against a healthy destination before it starts returning a response. From the caller’s point of view, what comes back is a normal response rather than an error
- A routing DSL written in YAML and CEL: Rules such as “always send requests from this project to a specific model” can be written declaratively as configuration, without touching your application code
- Guardrails and an agent firewall: PII masking and content policies are applied before charges are incurred. Tool calls and MCP calls made by agents can also be put through prior review
- Request logs and dashboard: You can review the score, the chosen model, latency and cost for each call in one place. Individual requests can be copied as cURL for reproduction and verification
Pricing
| Plan | Monthly price | What you get |
|---|---|---|
| Hacker | Free | 200+ model catalog, OpenAI-compatible API, automatic failover, basic dashboard, prompt versioning, 10 API keys, single-user workspace |
| Team | Contact for pricing | Everything in Hacker, plus up to 10 team seats, unlimited API keys, compliance enforcement and reports, priority support |
| Enterprise | Contact for pricing | Everything in Team, plus unlimited seats, dedicated infrastructure, a 99.99% uptime SLA, priority access to new models, dedicated support |
Token usage is billed separately from the plan fees above, at each provider’s published rates. The official site states that the routing fee is $0. Note that the amounts for the Team and Enterprise plans are shown as “Custom” on the pricing page, with no specific figures published. The FAQ states that the Team plan has no free trial and is billed from the first day.
Pricing is current as of September 2026. Please check the official site for the latest information.
Pros & Cons
✅ Pros
- You avoid holding contracts, SDKs and credentials for multiple providers, and your application side is done with a single OpenAI-compatible client
- Since no fee is added to token rates, you can adopt the gateway without worrying about extra cost created by the extra hop
- Even the free Hacker plan covers the model catalog and automatic failover, so you can check the behavior with real data before going to production
- When you want to change how requests are dispatched, a configuration change is enough — no redeploy of your application
- Because switchover during an outage is handled inside the gateway, one provider going down does not have to take your service down with it
⚠️ Cons
- The amounts for the Team and Enterprise plans are not published, so budgeting requires contacting the company
- There is no free trial, so trying the team-oriented features means paying from the start
- The hosted product itself is closed source; what is published is only “OrcaRouter-Lite”, a self-hosted version with a reduced feature set
- We could not confirm an official MCP server that external clients can connect to. On the MCP side, the emphasis is on features that review agent calls
- Having launched in May 2026, the service is still building up its track record for long-term stability and SLA performance
Comparison with Similar Services
| Service | How it differs from OrcaRouter |
|---|---|
| OpenRouter | Also a multi-model API router. OrcaRouter positions itself as an alternative and highlights zero-markup billing and score-based automatic routing as its differentiators |
| Portkey | A competitor strong in AI gateways and guardrails. The two overlap in enterprise governance and observability |
| LiteLLM | An open-source LLM proxy. It assumes you run it yourself, so its delivery model differs from OrcaRouter’s hosted SaaS |
| Martian | A service focused on model routing. OrcaRouter covers more ground by adding firewall and compliance features on top of routing |
Who Is It For
- Product teams that already use several LLM providers for different purposes and want to consolidate the call surface into one
- Developers who are sending even simple work to expensive models and want to revisit their cost breakdown
- Companies running agent and MCP tool calls in production that want auditing and restrictions applied before charges occur
- Anyone who wants to stabilize the backend of coding agents such as Claude Code, Codex or Cursor with failover in place
- People who want to start on the free tier and see how practical cross-model calling is in their own setup
Summary
The more model choices there are, the heavier the burden of deciding which to use when, and of keeping costs under control. OrcaRouter moves that decision into the gateway, and tries to do so without adding a fee to token charges. Since the core features are available on the free plan, a good starting point is to route only part of your existing traffic through it and see how the scoring and dispatch results look on the dashboard. Once you need team usage or compliance features, asking for a quote and deciding from there is the practical path.