Fly.io

What Fly.io actually is

Fly.io is a platform for running full-stack applications close to users. You hand it a Docker image or a project it can build into one, and it runs your code as small VMs in the regions you choose, behind a global Anycast network. Under the hood every workload is a Fly Machine — a Firecracker VM with a fast REST API — and on top of that sits Fly Launch, the orchestration layer most people actually use via flyctl and a fly.toml file.

The company is a Chicago-founded outfit (2016, co-founder and CEO Kurt Mackey) that raised through Y Combinator and has been through several pricing and platform reinventions since. That history matters when you read older reviews: the Fly.io of 2021 with its generous free allowances is not the Fly.io of today. New organisations now start on a free trial and then pay as they go, and a chunk of the product surface is optional paid extras — support plans, managed databases, Kubernetes.

What we like about the platform is its honesty of shape. There is no platform fee. There are no plans with arbitrary limits. You pay for Machines by the second while they run, for volumes by provisioned GB, and for data leaving the network by the GB. The pricing documentation publishes every billed resource with a number next to it, plus worked example bills. In our view that puts Fly.io in a small group of providers whose cost model you can actually reason about before deploying — see our /benchmarks page for how its smallest shapes compare on price.

The product surface spans three categories on ServerSpotter: /category/managed-cloud (Fly Launch plus Managed Postgres, Fly Kubernetes, support plans), /category/edge-computing (multi-region placement, Anycast routing, per-request routing with fly-replay), and the Machines API itself, which is effectively a small compute API for containers.

Pricing: what you really pay

Everything is metered, and the numbers are public. In the two cheapest regions (Ashburn and Secaucus), a started shared-cpu-1x Machine with 256MB RAM costs $2.19 per 30 days of continuous runtime; 512MB is $3.69 and 1GB is $6.70. A shared-cpu-2x with 2GB is $13.39. Dedicated CPU presets jump sharply: performance-1x with 2GB is $33.00 and performance-2x with 4GB is $66.00. Extra RAM beyond a preset is $6.00 per GB per month, and Machine prices vary by region — the same 1GB Machine runs $7.73 in Frankfurt and $8.50 in Singapore.

The metered extras are where bills grow. Volumes cost $0.15 per GB per month on provisioned capacity. Volume snapshots are $0.08 per GB per month with the first 10GB free. A stopped Machine still bills $0.15 per GB per month for its root filesystem. A dedicated IPv4 address is $2 per month (shared IPv4 and IPv6 are free). Outbound data is $0.02/GB in North America and Europe, $0.04/GB in Asia Pacific, Oceania and South America, and $0.12/GB in Africa and India; inbound is free.

Managed Postgres is the big-ticket item: $38 per month for Basic up to $1,922 for Performance, plus $0.28 per GB per month of storage. Fly Kubernetes is $75 per month per cluster before compute. Support is $29 per month Standard, $199 Premium, and Enterprise starts at $2,500. Fly's own worked examples put a small web app with a Basic Postgres database at $43.67 per month — of which the app Machine and its volume are $5.19. A 24/7 Discord bot is $2.21 per month. An app with one 1GB Machine each in Ashburn, Frankfurt and Singapore serving modest traffic lands around $27.20.

The free tier is gone. What exists is a trial: 2 VM hours or 7 days, whichever ends first, capped at 10 Machines, 20GB of volume storage, 2 vCPUs and 4GB RAM per Machine. Trial Machines stop automatically after five minutes of running. Adding a card ends the trial and starts billing. Our reading: Fly.io is cheap for always-on small compute and honest about it, but the database layer is priced like a managed service, not like the VM underneath it.

Performance & infrastructure

Fly runs its own hardware in 17 regions, and the current list is narrower than the "35+ cities" claim you will still find quoted around the web — regions have been consolidated over the past couple of years. Today you get Amsterdam, Stockholm, Paris, Dallas, Secaucus, Frankfurt, Ashburn, Johannesburg, Los Angeles, London, Tokyo, Chicago, Singapore, San Jose, Sydney, Toronto and Sao Paulo. That covers the US, Europe, East Asia, Oceania, South America and one African node; there is no India region despite older documentation listing one.

Region choice affects more than latency. Some regions carry WireGuard gateways for private networking, some can host Managed Postgres clusters, and capacity genuinely varies — Machine placement can fail in a full region, which is the kind of operational reality Fly's docs do warn about. Egress pricing also varies by continent, and Machine prices differ per region, so a global deployment is a cost calculation as much as a latency one.

On performance: these are Firecracker VMs that boot in well under a second from stopped state, and container fetch and rootfs assembly happens on first creation (low double-digit seconds). Auto-stop and auto-start can keep Machines idle at near-zero compute cost, with the stopped rootfs fee as the floor. Persistent storage is via Fly Volumes, which are tied to a single region and host. For object storage, Tigris is the native option — S3-compatible, provisionable with fly storage create, billed at the provider's list price through your Fly invoice. Upstash Redis and other extensions work the same way.

Reliability terms deserve a careful read. The published SLA commits to 99.9% platform uptime per calendar month — but only for Enterprise-plan customers, with service credits capped at 30% of the affected month's fees and a 30-day claim window. The definition of Unscheduled Downtime excludes hacking or virus attacks and reasonable emergency maintenance. For everyone else, there is no uptime commitment at all; the status page and the community forum are the recourse. Enterprise support response targets are concrete: 15 minutes for urgent outages, 4 hours for normal issues.

Where it falls short

The most common complaint we consider legitimate is cost surprise at the edges. Compute is cheap; the surrounding services are not. Managed Postgres at $38 per month is roughly the price of a capable VPS elsewhere, and a production three-node unmanaged Postgres cluster runs $82–164 per month in Fly's own estimate. If your mental model is "PaaS prices like a VPS", you will be disappointed once the bill arrives.

Second, this is a CLI-first platform. flyctl, fly.toml, fly-replay, WireGuard private networking, Machines versus Apps — the concepts are well documented but there is a learning curve before the mental model clicks, and the dashboard is thinner than what you get on Render or Heroku. Teams that want click-ops will fight it.

Third, the free tier exit stung existing users and will sting new ones: trial limits are tight, and hobby projects that used to coast on free allowances now cost real money. Combined with the $2 per month dedicated IPv4 charge and per-GB egress, a "free-looking" project needs a budget.

Fourth, platform risk. Fly's SLA carve-outs (hacking, virus attacks, emergency maintenance) and its Enterprise-only commitment mean smaller customers absorb the platform's bad days alone. Prices also vary by region and have been revised repeatedly — pin your assumptions to the current pricing docs, not to a two-year-old blog post.

Who should pick this

We would pick Fly.io for latency-sensitive APIs, globally distributed backends, real-time apps, and full-stack applications where running close to users matters more than squeezing the last euro out of compute. The single-Machine pattern — a small always-on service at $2.19–3.69 per month in a cheap region — is genuinely good value, and the three-region deployment pattern is affordable enough to be real rather than aspirational.

We would not pick it for cost-critical database-heavy workloads without doing the Postgres maths first, for teams with no CLI appetite, or for workloads that need a contractual uptime SLA without an Enterprise commitment. If you want maximum hardware per euro rather than a developer platform, our /benchmarks page will point you at raw providers instead.

Key takeaways

• Pay-as-you-go with no platform fee: smallest always-on Machine is $2.19 per 30 days (shared-cpu-1x, 256MB) in the cheapest regions; billed by the second.
• Every billed resource is published — volumes $0.15/GB/month, egress $0.02/GB in NA/EU, dedicated IPv4 $2/month, Postgres $38–$1,922/month.
• 17 regions on Fly's own hardware with Anycast routing — smaller than the historical "35+ cities" claim; prices and capacity vary by region.
• The free tier is gone: a 2-VM-hour/7-day trial is all new accounts get before pay-as-you-go.
• The 99.9% uptime SLA applies to Enterprise plans only; everyone else relies on the status page and community support.
• Best fit: latency-sensitive, globally distributed apps run by teams comfortable with a CLI-first workflow.

Sources

• https://docs.fly.io/about/pricing
• https://docs.fly.io/reference/regions
• https://docs.fly.io/machines/overview
• https://fly.io/legal/sla-uptime/
• https://fly.io/docs/about/free-trial
• https://fly.io/docs/tigris
• https://builtin.com/company/flyio
• https://techcrunch.com/2022/07/28/fly-io-wants-to-change-the-way-companies-deploy-apps-at-the-edge/