Neon is serverless Postgres done right. Storage and compute are separated so the database scales to zero when idle and back to full capacity in seconds, and branching is genuinely instant (copy-on-write at the storage layer, not pg_dump). The trade is that some traditional Postgres patterns assume always-on compute , connection pooling, long-running transactions, scheduled jobs. For per-PR preview databases, AI agent workflows, and elastic apps on Vercel or Cloudflare, Neon is the right Postgres.
- True scale-to-zero, pay nothing when idle, resume in seconds
- Instant branching via copy-on-write storage (not pg_dump)
- First-class Vercel + Cloudflare integrations
- Postgres 18 GA, native MCP server, Codex plugin published
- Acquired by Databricks 2025, well-capitalized, integrated with Databricks AI
- $19/mo Launch tier floor, free tier capped at 0.5 GB storage
- Scale-to-zero adds cold-start latency on first request
- Not the right fit for always-on heavy workloads (use traditional Postgres)
- Snapshot billing ($0.09/GB-month) adds up at scale
- Newer than RDS / Cloud SQL, less battle-tested ops tooling
Neon is a serverless Postgres platform with separated compute and storage. Spin up a database in seconds, branch it like a Git repo for previews and tests, and let it scale to zero when idle. Neon was acquired by Databricks in May 2025 and continues to ship independently, with deep integrations across the AI agent ecosystem (MCP, Codex, Cursor, Claude Code).
Where it fits
Neon is the right Postgres for the modern Vercel / Cloudflare app pattern: serverless functions hitting a serverless database, pay for what you use, scale to zero between bursts. The wedge is branching, every preview deploy gets its own real Postgres branch (copy-on-write, no data duplication), making per-PR isolation genuinely cheap.
It’s also a strong fit for AI agent workflows where short-lived agents need to provision real databases on demand, and for database-per-tenant architectures where the unit economics demand the database be idle most of the time.
For workloads that are always-on and traffic-heavy, a traditional managed Postgres (RDS, Cloud SQL, Supabase Pro) is often cheaper and simpler, Neon’s elasticity is worth real money only when compute genuinely sits idle.
Pricing in practice
Free includes 0.5 GB storage, 190 compute hours, and unlimited branches, fine for a side project, tight for anything real. Launch is $19/month with 10 GB storage, 300 compute hours, and PITR. Scale is $69/month with 50 GB storage and 750 compute hours. Business is $700/month for SOC2-bound workloads. Snapshot storage bills at $0.09/GB-month (started May 1 2026).
Compute is metered in CU-hours (1 CU = 1 vCPU + 4 GB RAM). The default auto-scaling range and the scale-to-zero timeout are both tunable per project, leaving them at defaults is the right move until you have real load data.
How it compares
Supabase, Managed Postgres with auth, storage, realtime, and edge functions bundled in. Pick when you want a complete backend in one workspace, not just a database.
PostgreSQL, Self-managed Postgres on your own infrastructure. Pick when you need full control over the runtime or when always-on workloads make serverless economics unattractive.
Convex, Reactive TS backend with end-to-end type safety. Pick when you want a TypeScript-native backend and don’t need Postgres compatibility.
What changed recently
Neon shipped neon.ts on 2026-06-15, a TypeScript
infrastructure-as-code surface that declares Postgres, Auth, and
Data API services alongside per-branch settings (TTL,
autoscaling limits, scale-to-zero) in a single file, applied via
neonctl config plan and neonctl config apply. The companion
@neondatabase/env package adds compile-checked environment
variable parsing so branch-keyed credentials fail at build time
rather than at first query. The release closes the gap between
Neon’s expanding service catalog and the project-level
configuration story, which previously relied on stitching CLI
commands together per service. Earlier in May 2026, Postgres 18
reached GA on Neon with a storage-layer rewrite delivering up to
5x write throughput, snapshot billing began at $0.09/GB-month
with new full_size and diff_size API fields, and Neon Auth
gained magic-link plus phone-number sign-in.
Sources
- Neon Changelog, neon.com, May 2026
- Neon Pricing, neon.com
- Databricks acquires Neon, databricks.com, May 2025
- Neon Docs, neon.com
- launch · 2026-07-14
Neon's new TypeScript SDK is built as much for agents as for people
Neon shipped @neon/sdk, a zero-dependency TypeScript client for its control-plane API that folds multi-step provisioning into single calls. The framing points at platform builders and automated callers, not just scripts written by hand.
- launch · 2026-06-16
PostgreSQL 19 Beta lands SQL/PGQ graph queries and concurrent table repacking
PostgreSQL 19 Beta 1 shipped on 2026-06-04 with SQL/PGQ property graph queries and a new REPACK ... CONCURRENTLY command, the largest categorical additions to vanilla Postgres since pgvector. GA is targeted at September or October 2026.
- launch · 2026-06-15
Neon ships neon.ts, bringing TypeScript infrastructure-as-code to a Postgres branching platform
Neon launched neon.ts on 2026-06-15, a TypeScript file that declares Postgres, Auth, and Data API services with per-branch compute settings. Applied via neonctl config plan and apply, the file gives Neon a Terraform-shaped workflow without the Terraform.
- beat · 2026-06-05
Supabase closes $500M Series F at $10.5B as AI agents take over Postgres provisioning
Supabase raised $500M at a $10.5B post-money valuation on June 4, 2026, led by GIC. The company says 60% of new databases on the platform are now provisioned by AI coding tools, with Claude Code the largest single driver and Multigres open-sourced alongside.
- comparison · 2026-05-26
Postgres in 2026: PostgreSQL, Neon, and Supabase compared
Postgres is the default database for new web applications in 2026. The harder question is which surface to run it on: raw PostgreSQL, Neon's serverless compute-storage split, or Supabase's open-source Firebase-alternative bundle. Here is how the three divide the work.