Neon announces AI Gateway, a new component of its serverless Postgres backend that routes LLM calls (both frontier and open-weight models) through a single Neon credential instead of separate API keys per lab. Hosted on Databricks' Foundation Model APIs infrastructure, it unifies billing on the Neon invoice with no markup over published lab pricing, supports OpenAI-compatible and native provider routes, and inherits Neon's branching model so each database branch gets its own gateway host and credentials. The gateway is currently in beta and free to use, with observability and spend controls planned as follow-ups.
Table of contents
What you get when model calls live in your backendNeon AI Gateway = Databricks scale and performanceSwitching overTry itQuestions this post answers
What is Neon AI Gateway and how does it work with the Neon Postgres backend?
Neon AI Gateway is a component of the Neon backend that lets you call LLMs using a single Neon credential instead of separate API keys per AI lab. It exposes an OpenAI-compatible /v1 path plus native routes for OpenAI Responses, Gemini, and Anthropic Messages, is hosted on Databricks' Foundation Model APIs infrastructure, and bills usage directly on your Neon invoice with no markup over published provider pricing. Developers wiring LLM calls into their backend can track releases like this one on daily.dev.
Do Neon AI Gateway credentials and endpoints work across database branches?
Yes, each Neon branch gets its own AI Gateway host and its own credential, following branch lineage rather than a single shared lab key. A credential created on main works on main and its descendants (preview, feature, CI branches) but not on sibling lineages, and deleting a branch removes its gateway host too, matching Neon's Postgres branching model. Teams evaluating branch-aware backend tooling can follow updates like this on daily.dev.
What performance benefits does Databricks' Foundation Model APIs infrastructure provide for open-weight models compared to standard hosting?
Open-weight models served through Databricks' Foundation Model APIs get the same serving optimizations as frontier models, including prompt caching, which delivers 2.5x higher throughput and 3x lower latency according to Databricks' own benchmarks. This infrastructure already processes more than 145 trillion tokens per month. Developers comparing inference infrastructure options can keep tabs on benchmarks like these via daily.dev.