The Fly.io in most reviews no longer exists

Most write-ups of Fly.io still describe three free 256MB VMs, 35-odd regions and GPUs on tap. None of that is what a new account gets. The free allowance was retired for new organizations on October 7, 2024. The region map has been consolidated to 18 locations (the docs table lists 17 codes), with 17 cities retired, including Madrid, Warsaw, Miami, Seattle, Montreal and Hong Kong. GPU Machines have been wound down: Fly published "We Were Wrong About GPUs" in February 2025, and customer notices quoted on Hacker News gave July 31, 2026 as the end date.

Then, on July 24, 2026, the company announced new funding (amount undisclosed), a new CEO in Scott Johnston as founder Kurt Mackey stepped aside, and a push behind Sprites, its sandboxed VMs for AI coding agents. The same post quotes Theo Browne saying Fly was the provider he was "least confident would be around by the end of the year." Fly's answer, in that post and again from staff on the forum, is that "Fly Machines and our Platform As A Service features aren't going anywhere."

So this review covers the product as it stands, checked in October 2026: pay-as-you-go only, a card on file, and a sharp price-per-second model with more operational work left to you than the marketing suggests.

What you are actually running

You hand Fly a Docker image and a fly.toml. Fly, in its own words, "transmute[s] containers into VMs": each instance is a Firecracker microVM called a Machine, running on Fly's own hardware and reachable through an Anycast network. Every app gets a shared IPv4 address and Anycast IPv6 at no charge. That is the genuinely unusual part. You get real VM isolation and a global edge without touching Kubernetes, and a Machine can stop when idle and cold-start on the next request.

Auto-stop is where the per-second billing earns its keep, since stopped or suspended Machines are not billed for CPU or RAM. The catch is in the docs, not the pitch: the proxy's stop loop "runs every few minutes and stops at most one Machine per region per pass," so a burst that scaled you to ten Machines takes a while to wind back down, and you pay for the tail.

Signup is quick, but the trial is short. New accounts get "2 hours of machine runtime or 7 days of access, whichever comes first," trial Machines stop after 5 minutes of running, and you are capped at 10 Machines and 2 vCPUs each. Adding a card ends the trial on the spot, and every organization needs one to keep going. There is no free hobby tier to park a side project on any more.

Pricing, checked in October 2026

Compute is billed by the second while a Machine runs. The pricing page quotes Ashburn (iad) by default, and other regions can differ.

Extra RAM is $5 per GB per month. Reserved blocks knock 40% off if you can commit for a year: $36/year buys $5/month of shared compute, $144/year buys $20/month of performance compute.

The line items that surprise people are the ones that keep ticking when nothing is running. Volumes cost $0.15/GB per month of provisioned size, attached or not, stopped Machine or not. A stopped Machine's root filesystem is billed at the same $0.15/GB. Volume snapshots have been billed since January 1, 2026 at $0.08/GB beyond the first 10GB. A dedicated IPv4 is $2/month, a static egress IP about $3.60/month, and certificates beyond the first ten are $0.10 each.

Egress is $0.02/GB in North America and Europe, $0.04/GB in Asia-Pacific and South America, and $0.12/GB in Africa and India. That last rate is six times the US price, which matters if Johannesburg was the reason you picked a global platform. Early-stage companies can apply for up to $15,000 in credits.

Volumes and Postgres: where the work lands on you

This is the section to read twice. A Fly Volume is "a slice of an NVMe drive on the same physical server as the Machine," it attaches to exactly one Machine, and "Fly.io does not automatically replicate data among the volumes." If that drive dies and you had one copy, the data is gone. Fly's own advice is to "always provision at least two volumes per app," which doubles the storage bill and leaves the replication logic to your application. Daily snapshots are kept for 5 days by default, and the docs say they "shouldn't be your primary backup method."

Managed Postgres (MPG) fixes much of that, at a price. Every plan includes high availability, automated backups and connection pooling:

• Basic, shared-2x with 1GB: $38/month
• Starter, shared-2x with 2GB: $72/month
• Launch, performance-2x with 8GB: $282/month
• Scale, performance-4x with 32GB: $962/month

Storage is $0.28/GB on top, up to 1TB per cluster. MPG runs in 13 of the 17 regions, and its own docs still list security patches and version upgrades, customer-facing alerting and migration tools as not yet available. Fly staff have said the first version's Kubernetes base "caused some stability problems"; the v2 rewrite was in beta in Dallas only as of May 2026. The old do-it-yourself Fly Postgres is still around from about $2/month, but it is listed as unsupported. For a $38 database without alerting, Render is the more finished managed option.

Support and reliability

Without a paid package, support is the community forum plus a billing email address. Email support starts at $29/month with a 36-hour first response. Premium is $199/month for 24 hours (1 hour on urgent issues) and a shared Slack. An uptime SLA exists only on Enterprise, from $2,500/month. For a platform whose volumes can lose data on a single drive failure, that is a thin safety net at the price most small teams will pay.

Reliability is the complaint that recurs across years. An HN user wrote in 2023 that their main complaint was "instability... My TLS cert did not renew, resulting in 8 hours of downtime," and forum threads carry titles like "Reliability: It's Not Great." It is not only history: Fly's status page logged 11 major or critical incidents between August 19 and September 29, 2026, including a critical network outage in Amsterdam on September 26 and Managed Postgres partially down in Chicago on August 20.

Billing is the other sore point. Fly's Trustpilot score sat at 2.8 across 20 reviews when checked, with users describing unexpected charges, being billed for idle database clusters and a dedicated IP added to a free account. Fly does run an accident-forgiveness policy and will talk refunds for mistakes or attack traffic through [email protected], but there is no spending cap to stop the bill in the first place.

Getting your data out

Leaving is easier than on most PaaS platforms, because what you deployed is a standard OCI image plus a config file. The image runs anywhere Docker does. Files on a volume come off with fly ssh sftp. For Managed Postgres, fly mpg proxy opens a local tunnel and plain pg_dump does the rest. Coming in is the harder direction: the MPG import docs say "there isn't currently an automated way to migrate an existing DB's data," so moving a live database onto Fly is a dump-and-restore window you schedule yourself.

Compliance is better than the company's size suggests: a SOC 2 Type 2 report, a DPA, and a pre-signed BAA through a $99/month HIPAA add-on available on any paid plan.

Verdict: who should use Fly.io, and who should skip it

Use Fly.io if you are comfortable with Docker and the command line, want your app running close to users on several continents, and have traffic that goes quiet enough for auto-stop to pay off. A small API on a shared-cpu-1x Machine for under $2/month, with real VM isolation and Anycast routing, has no close equivalent. It is also a natural fit if your product runs AI agents in sandboxes, which is where Fly is now putting its effort with Sprites.

Skip it if your app's value lives in a database you cannot afford to lose and you have nobody to run replication and backups. Skip it too if you want a human answer within a day without paying $29 a month for it, or if the July 2026 strategy shift makes you uneasy about a five-year commitment. Render gives you a more finished managed Postgres and a Heroku-style workflow. Railway is faster from git push to live URL. If you just want a server you control, a DigitalOcean Droplet is simpler to reason about and to bill.

Try Fly.io