Neon
Serverless Postgres with copy-on-write branching and 350ms cold starts, now Databricks-owned. Its Agent Plan exists because most new databases on it are created by agents.
Overview
Postgres with the storage layer rewritten underneath. Compute nodes are ephemeral and stateless; durable data lives in a separate multi-tenant service built from a pageserver and Paxos-replicated safekeepers. Two useful things fall out of that split. Idle projects scale to zero and resume in about 350ms, and branches are copy-on-write, so a full branch of a large database appears in seconds and costs nothing until it diverges from its parent.
The agent angle was not a marketing retrofit. By 2025 the majority of new database creations on the platform came from AI agents rather than humans, which is precisely what attracted Databricks; the deal was announced in May 2025 at a reported $1B against roughly $25M of ARR. Neon now sits inside Databricks as the Lakebase layer, and the Agent Plan targets platforms like app builders and coding agents that need to provision a database per end user with no human in the loop. Instant provisioning, unlimited projects, snapshots that version code and data as one unit.
Pricing after the acquisition is better than before it: storage at $0.35/GB-month against roughly $1.75 previously, compute at $0.106 per CU-hour on Launch and $0.222 on Scale, and a free tier of 100 CU-hours and 0.5 GB per project with 10 branches. Neon Auth and a REST Data API come bundled. The engine is Apache 2.0 on GitHub at around 23k stars and 8,400-plus commits, but the control plane that actually delivers branching and autoscaling is not published, so read the open-source label as partial rather than as an exit route.
Key Features
- ✓ Storage separated from compute, so an idle project scales to zero and wakes in roughly 350ms
- ✓ Copy-on-write branching: forking a large database is instant and costs nothing until the branch diverges
- ✓ Agent Plan with unlimited projects, aimed at app builders that hand every end user their own database
- ✓ Snapshots that checkpoint code and data together, plus point-in-time restore
- ✓ Standard Postgres extension set including pgvector, so ORMs, migration tools and psql all behave normally
- ✓ Apache 2.0 storage engine (pageserver, Paxos-replicated safekeepers, compute node) on GitHub at around 23k stars
Where it holds
- • Real Postgres rather than a wire-compatible imitation, which means existing extensions, drivers and migration tooling work without adaptation
- • Branching turns preview environments and agent checkpoints into a cheap primitive instead of a nightly restore job
- • Storage dropped to $0.35/GB-month after the Databricks deal, roughly a fifth of the earlier rate
- • Scale to zero makes a fleet of thousands of mostly-idle per-tenant databases affordable, which is the actual agent-platform shape
Where it breaks
- • Databricks bought it in May 2025 for about $1B and folded it into Lakebase. Roadmap control now sits with a much larger company with its own priorities.
- • Only the storage engine is Apache 2.0. The control plane, autoscaling and branching orchestration are not in the repo, so self-hosting a true Neon is not realistic.
- • A 350ms cold start is fine for a web request and irritating in a chatty agent loop that reconnects constantly
- • CU-hour billing is hard to forecast. Autoscaling means one bad query plan can quietly multiply the compute line.
My Take
Branching is the feature that matters, and it is genuinely different from a snapshot: copy-on-write, so forking a 50 GB database is instant and free until something writes to it. That maps onto agent work almost too neatly, which is why the company reported over 80% of new databases being created by agents rather than people, and why an Agent Plan now exists at all. Weigh the acquisition properly though. Databricks paid around $1B in May 2025 against roughly $25M of ARR, and while storage prices fell hard afterwards to $0.35 per GB-month, a small team building on it is betting on someone else's platform strategy. If you want auth, storage and edge functions in the same box rather than just the database, Supabase is still the better fit.
Quick Info
- Pricing:
- freemium
- Openness:
- Open source
- Licence:
- Apache 2.0
- Starting at:
- Free gives 100 CU-hours of compute and 0.5 GB of storage per project with 10 branches. Launch is pay as you go at $0.106 per CU-hour, Scale at $0.222 per CU-hour, both with storage at $0.35/GB-month, down from roughly $1.75 before the Databricks acquisition. Branches beyond the 10 included on Launch or 25 on Scale bill at $1.50 per branch-month, prorated hourly. Neon Auth and the Data API are bundled at no extra charge. A separate Agent Plan covers platforms provisioning databases for their own end users: unlimited projects at Launch compute rates plus credits toward a free tier. Enterprise is quoted.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Neon holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Neon
Convex
freemiumReactive TypeScript backend where queries, mutations and durable workflows run inside the database. Agent and Workflow components handle long-running tool loops.
TiDB
open-sourcePingCAP's MySQL-compatible distributed SQL database, Apache 2.0 across TiDB, TiKV and TiFlash. Vector columns and HNSW indexes ride on the TiFlash columnar replica, next to transactional rows.