Neon Review

Neon is the Postgres I would attach to a preview app, and the one I would not hand a finance team that wants a ceiling the platform enforces. Launch and Scale bill what ran, which fits a database that sits idle most of the day and gets expensive once you force it to stay awake.

A compute unit is about 4 GB of RAM, so that hourly rate buys a larger machine than an Aurora ACU. Storage, history, and extra branches keep billing after compute has suspended, which is what a sleeping database still costs.

Toolradar data: the September 2026 managed databases ranking evaluated 23 tools.

Its buyer note names Neon, Supabase, and PlanetScale beside the hyperscalers.

Railway and Render have directory pages too.

How we compared: each price was read on the vendor's own page on September 24, 2026. Neon was checked on the pricing page and the plans document. Aurora figures are Amazon's US East (Northern Virginia) examples. PlanetScale figures are the AWS us-east-1 catalog.

Neon is Databricks' serverless Postgres, sold on the pricing page as Lakebase Postgres: a project, branches, and compute that can suspend. It speaks ordinary Postgres, so your ORM still works and a dump is still the exit if you need to leave.

Branching, the pooler, autoscaling, and extensions such as pgvector, PostGIS, and TimescaleDB are on every plan, including Free, so a side project can test a vector column or a second branch without upgrading. Auth, object storage, and functions are still marked beta, so do not leave a working backend to chase them.

Managed Better Auth caps Free at 60,000 monthly active users, enough for a demo and the wrong cap for a consumer app. Storage and functions are not charged during beta, so they are not the invoice to forecast.

If the job is auth plus files plus a database, that is a backend-as-a-service decision, and Supabase is the closer product.

You create a project, which arrives with a root branch. Child branches are copy-on-write: storage stays at zero until that branch is written, then bills those writes, capped at the logical size of the data.

Compute is CU-hours, size times hours running, so a 0.25 CU box for four hours is one CU-hour and a suspended compute adds nothing. Sleeping previews stay cheap, and warm production databases do not.

On Free you cannot turn suspend off, and the default is five idle minutes. Neon says the next query wakes the compute in a few hundred milliseconds, which a preview can absorb and a user-facing request should not.

Launch can disable suspend, and Scale can set the window from one minute through always on.

Computes larger than 16 CU never suspend, and Scale's fixed sizes go to 56 CU, so those sizes never get the idle rate. Logical replication keeps compute active while a subscriber is connected, so that pipeline stays on the meter and a sync can erase the savings.

A read replica is a second compute bill, not a free feature. The database clients list is the SQL tool you keep beside this.

Neon publishes no seat price on Launch or Scale, so headcount is not the meter. You pay CU-hours while compute is running, then storage, history, and branches past the included count. Invoices under $0.50 are not collected, so a tiny preview may never charge the card.

Neon's own light Launch example totals $2.31: about 10 CU-hours, 2 GB on the root, 1 GB on a child branch, and 1 GB of history. That total is the bill when the database mostly sleeps. Their Scale example, 4 CU for roughly 17 billable days plus storage and history, is $431.15 for a database left warm for most of the month.

Free is an allowance, then a wall, not a plan that starts charging. Past the compute allowance or 5 GB of egress, compute suspends until the next billing period, so the demo goes dark instead of invoicing you. Past the storage allowance, writes block and the data is not deleted.

Paid egress includes 500 GB per project before the table's per-GB rate, and a branch can grow to 16 TB.

Supabase Pro is an organization fee, and the compute credit covers one Micro, so a second always-on project is another line. Spend caps are on by default there, which is the brake Neon does not offer.

Aurora Standard in US East (Northern Virginia) bills $0.12 per ACU-hour, and an ACU is about 2 GiB, half the memory of a Neon CU. PlanetScale's cheapest single-node Postgres SKU in AWS us-east-1 is a monthly cluster, paid even when idle.

Railway does not bill CPU while a service is stopped, though the volume keeps charging. Render sells a named size that stays on, so you pay for the reserved month rather than hours queried.

The longer host comparison is Postgres hosting, and if the app host is the purchase, use serverless platforms and the Vercel review.

Neon fits a developer who wants a branch per pull request and can tolerate a few hundred milliseconds when a preview wakes. Side projects fit Free until they outgrow the allowance and have to move. Launch fits a startup whose traffic is bursty enough that the database still sleeps.

It is the wrong default when the database must stay warm, because disabling suspend is a Launch switch and a large compute cannot suspend at all. It is also the wrong default when a compliance report has to be in hand before Scale, or when auth and file storage are the product rather than a beta beside Postgres.

Skip Neon when you need a cap that refuses traffic. Launch and Scale email you at 80% and 100% of a threshold you set, and the query still runs, so the alert does not stop the bill.

Supabase's spend cap is the closer brake, and a phone client that needs offline sync belongs on the Firebase review, not on a Postgres branch.

A VM you patch yourself is a different invoice, covered in the DigitalOcean review. The Supabase vs Firebase note is the backend fork if Neon was only in the running because it said Postgres.

Devshot data: this newsletter goes to 27,887 developers, and the audience file updated September 20, 2026 shows a 30.3% average open rate. Subscribe free if you want the next meter change in the inbox.