Is OpenRouter safe for production? Data, fees and limits to check
What to check before real customer traffic goes through it.
Quick answer
OpenRouter is widely used and fine for production if you set it up deliberately. Require zero data retention providers if your prompts are sensitive, put a credit limit or guardrail on every key, restrict which models and providers each key can use, and price in the platform fee (5.5% on Standard). If you need an IP allowlist without an Enterprise plan, or one place for keys, caps and cost proof across providers, put a gateway like Prune in front.
1. Where your prompts go
OpenRouter is a router. Your request is served by a third-party provider (OpenAI, Anthropic, Google, or one of many hosts for open models), and each one has its own logging and retention policy. OpenRouter's own terms say it does not guarantee those providers meet your security requirements. That is normal for a router, but it means "is OpenRouter safe" is really "are the providers my requests land on safe for this data".
- Require zero data retention (ZDR) per request or on a guardrail so only ZDR endpoints are used.
- Restrict allowed providers on a guardrail so a cheap unknown host is never picked for a sensitive call.
- Keep personal data out of prompts you send to providers you have not reviewed.
2. Spend limits per key
This is an area where OpenRouter is ahead of going direct: OpenAI and Anthropic have no per-key limit (see OpenAI API key spending limit). OpenRouter keys can carry a credit limit, and guardrails add a dollar limit with a reset interval plus allowed models. Use one key per service, set a limit on each, and check remaining credit with GET /api/v1/key.
3. Fees
Model prices are passed through, and a platform fee is added: 5.5% on Standard, 8% on Business. With your own provider keys (BYOK), the first $25,000 of list-price inference a month is free, then 5%. The free tier allows 50 requests a day on free models, which is for testing, not production.
4. Reliability
Routing across providers is a reliability feature: if one host is down, another can serve the same open model. The flip side is that two calls to the same model can land on different hosts with different speed, context limits and quantization. Pin providers for workloads where consistent output matters, and log which provider served each call.
5. Key security
An OpenRouter key can reach every model you have not blocked, so a leaked one is expensive. Keep it server side, set a credit limit, and restrict models. IP allowlisting is on the Enterprise plan. If a key leaks, the steps in API key leaked apply here too.
Production checklist
- One key per service, each with a credit limit.
- A guardrail with allowed models, allowed providers and a reset interval.
- ZDR required for anything with customer data.
- Provider pinned where output consistency matters.
- The platform fee included in your unit economics.
- Alerts on unusual spend, not just a monthly review.
Using OpenRouter with Prune
Prune is an AI gateway, and OpenRouter can be one of its providers. Your OpenRouter key goes into Prune's encrypted vault, and your app holds only a Prune key. Shield adds a daily spend cap, spike detection, an IP allowlist and allowed models on each Prune key without an Enterprise plan. Free OpenRouter models are never swapped for paid ones, and every call gets a signed receipt showing what was served and what it cost. You can also send some traffic to OpenAI, Anthropic or Gemini directly and keep one dashboard for all of it.
With a single line of code, Prune cuts your AI bill and protects your team from runaway spend, leaked keys and dangerous calls.
