Netlify's AI Gateway Adds OpenRouter, but Only Privacy-Safe Models

Netlify’s AI Gateway now reaches models it doesn’t integrate directly, by routing requests through OpenRouter to providers including DeepSeek, Meta, Mistral, Qwen and xAI. Only models whose provider guarantees a Zero Data Retention (ZDR) policy at request time are reachable, so a chunk of OpenRouter’s catalog stays off-limits (Netlify changelog, Aug 6 2026).

What shipped

AI Gateway launched in beta on Oct 1, 2025 alongside Agent Runners, with direct integrations to OpenAI, Anthropic and Google Gemini (Netlify changelog). Those three providers are still served through Netlify’s own API calls to each vendor. The new piece is OpenRouter as a fourth path: instead of Netlify building and maintaining separate integrations for every additional model provider, it hands off to OpenRouter’s routing layer to reach DeepSeek, Meta, Mistral, Qwen and xAI models.

Netlify’s filter sits on top of that hand-off. Only OpenRouter models with at least one upstream provider currently offering a ZDR guarantee, meaning prompts and outputs aren’t stored or used for training, show up as available through the gateway. A model on OpenRouter without a ZDR-compliant provider at request time simply doesn’t appear. Usage is billed through existing Netlify credits, with no separate OpenRouter account or API key to manage; the gateway injects OPENROUTER_API_KEY and OPENROUTER_BASE_URL into Functions, Edge Functions and the Preview Server automatically.

Why the gate matters more than the roster

Model choice is the headline. The constraint Netlify put on that choice is what developers should actually weigh. OpenRouter’s whole pitch is breadth, one API key to more than 400 models across 70-plus providers (Vercel AI Gateway vs. OpenRouter comparison), and a chunk of that breadth comes from providers who don’t offer a no-retention guarantee. By gating on ZDR, Netlify is trading catalog size for a privacy floor it can stand behind: developers get access to open-weight and non-Big-3 models, but not the full OpenRouter surface, and which models qualify can shift over time as providers change their retention terms.

That’s a different bet than Vercel’s. Vercel’s own AI Gateway, which predates this Netlify move, already covers 200-plus models across 40-plus providers direct, at provider list price with no token markup (Vercel AI Gateway vs. OpenRouter comparison). Netlify isn’t matching that breadth with direct integrations; it’s borrowing OpenRouter’s aggregation and then narrowing it. For a team picking a platform partly on AI model access, that difference plays out concretely: Vercel’s catalog is wide by default, Netlify’s is wide only within the ZDR subset.

Our take: gating on Zero Data Retention is the right instinct for a platform whose customers are shipping other people’s production apps through it, an agent that leaks a customer’s prompts to an undisclosed training pipeline is a worse outcome than an agent with fewer model choices. But it also means Netlify’s “OpenRouter support” headline undersells the constraint, developers should check which models currently qualify before building a workflow around one, since ZDR status is a provider-side policy Netlify doesn’t control and can change without a Netlify-side changelog entry.