Neon Blog

Neon's AI Gateway now exposes POST /v1/embeddings on the same OpenAI-compatible /v1 path and credential used for chat, accepting a single string or a batch of up to 150 strings, with qwen3-embedding-0-6b (1024 dimensions, configurable) and gte-large-en (1024 dimensions, cosine distance) available at launch. Vectors can be written into Lakebase Postgres and queried with Lakebase Search, keeping the full retrieval pipeline inside one Neon branch.

Neon Free plan projects now get 1 GB of Postgres storage each, up from 0.5 GB, with existing projects updated automatically and the 100-project limit unchanged. Each project also includes 100 CU-hours of compute per month, autoscaling to 2 CU, 10 branches, a 6-hour instant restore window, Managed Better Auth for up to 60,000 MAU, 5 GB of S3-compatible object storage, and Node.js functions with 1M monthly invocations.

Neon outlines reliability practices adopted after two control-plane outages in one week, including postmortems that lead with customer impact and quantify detection and mitigation times, plus a rule to halve each response interval. The post also covers AI-assisted impact measurement, where agents must produce rerunnable notebooks pairing each figure with its query.

Neon details how Lakebase Postgres's decoupled compute-and-storage architecture makes restores a metadata operation that creates a branch at a timestamp, dropping recovery time to seconds even for 100 TB databases instead of the multi-hour provisioning, snapshot hydration, and WAL replay of traditional managed OLTP. The post cites a survey of 50 developers running production Postgres in which 59% had a critical production failure in the past 12 months and only 21% recovered in under 60 minutes.

A new storage_object_created trigger invokes a Neon Function when an object is created in an Object Storage bucket, passing the bucket name and object key in the POST body and supporting an optional key prefix to scope it to a path. Functions are long-running and remain compatible with scale to zero, and the schedule trigger is also available.

Function Triggers let a branch-scoped cron schedule invoke a deployed Neon Function, keeping the timer outside the Postgres compute so scale to zero can stay on, unlike pg_cron. Triggers run JavaScript or TypeScript that can call HTTP, Object Storage, and the AI Gateway, and child branches inherit parent triggers disabled until enabled.

Claimable Neon lets AI agents provision temporary Neon projects using anonymous registration without creating an account, then humans can sign in later and claim the project into their organization. The feature implements the auth.md protocol, supports Postgres, Data API, and Managed Better Auth, and uses 72-hour expiration with 100 MB storage and 1 GB transfer caps for unclaimed projects.

Neon now sizes Postgres shared_buffers to 75% of the compute instance's memory capacity automatically, making the default more robust than a fixed 128 MB. The capacity-aware default applies across compute profilesholones; for most workloads it requires no configuration, though specific tuning guidance applies to multi-instance deployments.

A new @neon/tools package wraps the @neon/sdk with Zod-validated, typed agent tools, shipping adapters for MCP, Mastra, and Eve, plus a CLI to run a Neon MCP server.

Blog post describes how Lakebase Postgres uses autoscaling and scale to zero to avoid paying for idle Postgres instance hours, and surveys patterns this unlocks.

Neon is now available as a plugin in Grok Bot. After connecting your Neon account once, you can ask a Grok bot to build on Neon.

Technical guide explains how Neon's Lakebase Postgres autoscaling determines when to resize compute by tracking three signals: CPU load, memory use, and compute-cache working set size. The architecture separates compute and storage layers, enabling in-place VM resizing without stopping the database, and uses HyperLogLog cardinality estimation to measure the workload's active data footprint.

An experimental space for Postgres tooling