To try an open-weight model safely, put three controls in place first: a data policy your coding agents can't route around, a provider boundary only team owners can change, and a rollback that is one line in the agent's model picker. Vercel AI Gateway ships all three as team-wide settings.
Copy link to headingWhat are open-weight models, and why do teams run them?
Open-weight models publish their weights, so multiple providers can host and serve the same model. When people ask about trying "the latest open-weight release," they usually mean models like DeepSeek, Kimi (from Moonshot), or GLM (from Z.ai). Those names match how the AI Gateway Production Index measures the category: DeepSeek, MiniMax, Moonshot, and Z.ai serving their own models.
Teams are not waiting for permission to run them. In the August 2026 AI Gateway Production Index, open-weight models grew from 11% of gateway volume in April to 36% in July, priced at about a seventh of frontier rates. That growth and that price gap explain why someone on your team wants to try this week's release.
The sticking point is usually governance rather than model quality. Where does your data go, who caps the spend, and how do you undo it if the model disappoints? Each of the three controls in this guide answers one of those questions.
Copy link to headingSet a data policy agents can't route around
Zero Data Retention (ZDR) in Vercel AI Gateway is a team-wide policy. Enable it once from the AI Gateway dashboard settings, and every request routes only to providers that are ZDR-compliant through Vercel's agreements. It's available on Pro and Enterprise plans and costs $0.10 per 1,000 requests, charged only on successful responses.
Team-wide is the point. The policy applies to every request without touching any agent's configuration, so a coding agent can't route around it, and there is no per-agent setting to forget.
Two defaults keep the policy conservative:
Providers whose retention stance Vercel doesn't know, or has no agreement with, are treated as non-ZDR-compliant. They are not routed to while ZDR is on.
ZDR is a superset of disallowing training: every ZDR-compliant provider also disallows training on your prompts.
Vercel's own gateway holds itself to the same standard. It does not retain prompts or outputs, and user data is deleted after requests complete. The Zero Data Retention docs cover the full policy.
Copy link to headingDraw a provider boundary only team owners can change
The provider allowlist lets team owners restrict which providers can serve requests through the gateway. It's opt-in and available on Pro and Enterprise plans at $0.10 per 1,000 successful requests; blocked requests are not charged.
Turning it on adds every current provider to the allowlist, so existing traffic never breaks. From there, you remove the providers you want to block. Disabling a provider removes it from routing for all requests, and automatic fallbacks apply only to providers that are still allowed. New providers added to the gateway later start disabled until an owner enables them.
Like ZDR, the allowlist applies to every request from the team, and only team owners can change it. If no allowed provider can serve a request, the request is refused. The provider allowlist docs describe the full behavior.
Copy link to headingKeep the rollback to one line
You need a Vercel account and the Vercel CLI. From there, vercel ai-gateway coding-agents setup is the one-line setup for 9 supported coding agents, and any of 200+ models works in any agent.
Switching models is a one-line change in the agent's model picker (for example, /model). If the new open-weight model disappoints, picking your previous model again is the whole rollback.
For the spend side, budgets are available for the keys your agents use; setting them up is covered in the guide to using any coding agent with 300+ models.
This is the general shape of a safe trial anywhere: a data policy individuals can't override, an explicit boundary on who serves the traffic, a spend cap, and a rollback cheap enough that trying costs nothing to undo. The gateway's team-wide settings are one implementation of that practice.
Copy link to headingWhat these controls are not
They are not free add-ons. ZDR and the provider allowlist are Pro and Enterprise features, each priced at $0.10 per 1,000 requests. ZDR is charged per request on successful responses; the allowlist is charged per successful request, and blocked requests are not charged.
ZDR is not silent when a model doesn't qualify. If no ZDR-compliant provider is available for a requested model, the request fails rather than routing to a non-compliant provider. That is the policy working, but plan for it. A model served only by non-compliant providers is off the table while ZDR is on.
This page is not a model verdict. The Index numbers describe usage and price, not a ranking. Which model deserves your subscription money is a separate decision; see the related resources below.
Copy link to headingFrequently asked questions
Copy link to headingDoes Zero Data Retention apply to every request my team makes?
Yes. Zero Data Retention in Vercel AI Gateway is enabled team-wide from the AI Gateway dashboard settings, and while it is on, every request your team makes goes only to providers that are ZDR-compliant under Vercel's agreements. The policy sits at the gateway rather than in any agent's configuration.
Copy link to headingCan a coding agent bypass the provider allowlist?
No. The provider allowlist in Vercel AI Gateway covers every request the team sends, and editing it requires the team owner role. Automatic fallbacks also stay on the allowed list, and when none of the allowed providers can serve a request, the gateway refuses it.
Copy link to headingWhat happens if a model has no Zero Data Retention provider?
The request fails instead of routing to a non-compliant provider. While Zero Data Retention is on, Vercel AI Gateway counts any provider whose retention stance it doesn't know or doesn't agree with as not ZDR-compliant. A failure here means the policy is held, so pick a model that a ZDR-compliant provider serves.
Copy link to headingHow do I go back to my old model if a new one disappoints?
Pick your previous model in your coding agent's model picker (for example, /model). A model switch through Vercel AI Gateway is a one-line change in the picker, so rolling back takes the same single step as trying the new model did.
Copy link to headingDoes Zero Data Retention also stop providers from training on my prompts?
Yes. Zero Data Retention is a superset of disallowing prompt training, so every ZDR-compliant provider Vercel routes to also disallows training on your prompts. The gateway itself retains no prompts or outputs.
Build a software factory with Vercel Connect
Ship a software factory with GitHub and Linear connectors provisioned for you from the first deployment.
Deploy now