{"id":8838,"date":"2026-10-08T04:38:41","date_gmt":"2026-10-08T04:38:41","guid":{"rendered":"https:\/\/suprmind.ai\/hub\/?p=8838"},"modified":"2026-10-08T04:42:09","modified_gmt":"2026-10-08T04:42:09","slug":"is-openrouter-safe","status":"publish","type":"post","link":"https:\/\/suprmind.ai\/hub\/insights\/is-openrouter-safe\/","title":{"rendered":"Is OpenRouter Safe? What Actually Happens to Your Prompts"},"content":{"rendered":"\n

OpenRouter routes API calls to more than 400 models from dozens of providers through one endpoint, which makes it a popular default for anyone who doesn’t want to manage separate accounts with OpenAI, Anthropic, Google, and everyone else. <\/p>\n\n\n\n

That convenience comes with a tradeoff worth understanding before you send anything sensitive through it: every request now crosses OpenRouter’s own infrastructure and<\/em> whichever downstream provider it routes to, and the rules governing each layer aren’t identical.<\/p>\n\n\n\n

Here’s what OpenRouter’s own documentation, its terms of service, and user reports actually say about what happens to your data.<\/p>\n\n\n\n

What OpenRouter stores by default<\/h2>\n\n\n\n

By OpenRouter’s own documentation, prompt and response content is not stored unless you explicitly opt in. What it does store regardless of your settings is request metadata – token counts, latency, model used, timestamps – retained for billing and reporting. <\/p>\n\n\n\n

OpenRouter also samples a small number of prompts for anonymous categorization to power its model-ranking reports; if you haven’t opted into anything else, that sample isn’t tied to your account.<\/p>\n\n\n\n

So out of the box, the defaults are reasonably private. The risk shows up in what happens when you turn features on.<\/p>\n\n\n\n

The two opt-ins, and why they’re not the same thing<\/h2>\n\n\n\n

OpenRouter has two separate settings that expose prompt content, and it’s easy to conflate them:<\/p>\n\n\n\n

    \n
  • Private Input & Output Logging<\/strong> (Observability settings): stores your prompts and completions so you can review them later for debugging. OpenRouter says it doesn’t access this data itself, and for organizations only admins can view it. It’s stored in an isolated cloud storage bucket, encrypted at rest, and retained for a minimum of three months – possibly longer, at OpenRouter’s discretion, unless you request deletion.
    <\/li>\n\n\n\n
  • OpenRouter Use of Inputs\/Outputs<\/strong> (Privacy settings): a separate opt-in that lets OpenRouter use your prompt and completion data to improve its product, in exchange for a 1% discount on usage. This is the setting that actually grants OpenRouter rights over your content, not just storage.<\/li>\n<\/ul>\n\n\n\n

    Both are off by default. The distinction matters because the second one is where the real data-rights question lives – and it got more explicit recently.<\/p>\n\n\n\n

    A quiet terms update worth knowing about<\/h2>\n\n\n\n

    OpenRouter’s terms of service were updated in July 2026. According to third-party terms-tracking analysis, the revision clarified that opting into prompt logging now automatically enables chat logging as well, and that doing so grants OpenRouter a broad, perpetual license to use that content for commercial purposes – including the right to license or sell it in anonymized form. <\/p>\n\n\n\n

    The setting itself is still opt-in and still reversible going forward, but the rights granted while it’s on are not retroactively undone by turning it back off.<\/p>\n\n\n\n

    For anyone routing confidential code, client data, or regulated information through OpenRouter, this is the setting to leave alone.<\/p>\n\n\n\n

    The two-hop problem<\/h2>\n\n\n\n

    Even with prompt logging fully off, your data still makes two administrative stops: OpenRouter, then whichever provider it routes your request to. OpenRouter’s own retention policy doesn’t change what the downstream provider does with the same prompt – some providers train on API data by default, some don’t, and OpenRouter documents each provider’s policy per model so you can filter for ones that don’t retain or train. <\/p>\n\n\n\n

    OpenRouter also offers Zero Data Retention (ZDR) controls, enforceable globally, per model group, or per individual request, which restrict routing to only ZDR-compliant endpoints. One limitation: ZDR enforcement covers inference routing, not plugins or tools you enable, like web search – those may run through third-party services with their own retention rules.<\/p>\n\n\n\n

    What users actually report<\/h2>\n\n\n\n

    Public reviews and forum threads add a second dimension beyond the written policy: account and billing stability. Recurring complaints on Trustpilot and Reddit describe accounts flagged and locked for “suspicious activity” after normal use, purchased credits that failed to appear or vanished from the dashboard, and support tickets on billing discrepancies going unanswered for weeks. <\/p>\n\n\n\n

    OpenRouter’s terms give users only a 24-hour window to request a refund on unused credits, credits expire after 365 days, and cryptocurrency payments are non-refundable under any circumstances. None of this makes OpenRouter unsafe to use for casual or low-stakes work – plenty of Reddit threads describe it working fine for exactly that – but it’s a different risk profile than a direct provider relationship, and worth weighing if you’re routing anything business-critical through a prepaid balance.<\/p>\n\n\n\n

    <\/p>\n\n\n\n

    \n