Fly.io pricing in 2026: what the $5 plan everyone quotes actually costs now

Search "fly.io pricing" today and you get a museum of dead numbers. Half the results quote a $5/month Hobby plan and a $29 Launch plan. Those are the 2024 legacy tiers. They are gone for new accounts. This teardown decodes what a Fly.io app actually costs in 2026, where the meter hides, and three real 30-day bills.

_Correction, October 11, 2026: every compute figure on this page has been re-based against Fly.io's current published rates, re-verified on this date against its resource pricing docs and its pricing page{target=blank}. This page previously published $2.02 a month for a shared-cpu-1x 256MB Machine, $32.19 for a performance-1x 2GB Machine, about $801 for a performance-16x 32GB Machine, $5 per 30 days per additional GB of RAM, and per-second rates of $0.00000078 and $0.00001242. The correct figures are $2.19, $33.00, $528.02, $6.00, and $0.0000008465 and $0.000012732. All three worked bills have been recomputed, and the reserved variant of bill B, which previously read $18.60 of compute and a $52.10 total, is corrected below. The fuller and continuously maintained version of this teardown lives at Fly.io Pricing in 2026: The Real 30-Day Bill, which carries the full preset ladder, the 17-region multiplier table, and the meters this page does not cover: cross-region private network transfer billed in both directions, Managed Postgres as a cluster plan from $38 a month, Sprites, Fly Kubernetes and static egress IPs.

Quick Answer (2026)

Fly.io in 2026 is pure pay-as-you-go: you pay per running second of a Machine, plus storage, bandwidth, and an optional support plan. A shared-cpu-1x 256MB machine running 24/7 is $2.19/month of compute in Fly's two lowest-priced regions. A small production performance-1x 2GB machine is $33.00/month of compute. There is no free tier for new customers anymore, and the widely quoted "$29/month plan" is a support tier, not a hosting plan. Reserve steady machines for 40% off; leave spiky ones on-demand.

What actually changed

Fly.io retired named subscription plans and moved fully to usage billing. The old free allowance (up to 3 shared-cpu-1x 256MB machines and 3GB of volume storage) is now marked legacy and applies only to grandfathered accounts. A new account pays from the first second a Machine is running. That single change is why so much of the internet's Fly.io advice is now wrong.

Here is the gap between what search results still say and what the meter charges in 2026.

The real 2026 meter, decoded

Compute is billed per second while a Machine is started. Converting the per-second rates to a 24/7 month (about 2,592,000 seconds) makes them legible:

Those are the rates in Ashburn (iad) and Secaucus (ewr), Fly's two baseline regions at a multiplier of 1.000. Every other region carries a published multiplier, from 1.038 in Amsterdam and Stockholm up to 1.615 in Sao Paulo, so the same shared-cpu-1x 256MB Machine is $2.28 in Amsterdam and $3.54 in Sao Paulo.

Extra RAM is a separate line: $6.00 per 30 days per additional GB in the baseline regions, on top of whatever the preset already includes. Then the storage, bandwidth and support lines a small app actually hits:

• Volumes: $0.15/GB per month of provisioned capacity (you pay for what you allocated, not what you filled).
• Volume snapshots: $0.08/GB per month, first 10GB free each month.
• Stopped-machine storage: about $0.15 per GB of rootfs for a machine stopped 30 days.
• Bandwidth (egress): $0.02/GB in North America and Europe, $0.04/GB in Asia Pacific, Oceania and South America, $0.12/GB in Africa and India.
• Dedicated IPv4: $2/month.
• Support (optional): Standard $29/mo, Premium $199/mo, Enterprise from $2,500/mo. HIPAA workloads add $99/mo.

All figures above are from Fly.io's own resource pricing docs and pricing page, re-verified against both on October 11, 2026.

Trap 1: scale-to-zero is not scale-to-free

Fly Machines can auto-stop when idle, and teams treat that as a $0 state. It is not. A stopped machine bills nothing for compute, but any attached volume keeps billing at $0.15/GB per month of provisioned size, and the stopped rootfs still costs about $0.15/GB for 30 days.

A "scaled to zero" service with a 3GB volume is not $0. It is about $0.45 to $0.60/month of storage doing nothing. Multiply that across a dozen sleeping preview environments and the idle bill is real.

Trap 2: the $29 is support, not the server

On a small Fly.io bill, the largest single line is often the support plan, not the compute. Standard support is $29/month, which is more than thirteen shared-cpu-1x machines running 24/7. Support is genuinely useful once you are in production, but pricing it as if it were the hosting cost is how people conclude "Fly.io is expensive" when the compute underneath is a couple of dollars.

The lever: reserve steady machines for 40% off

Fly.io gives a 40% discount when you reserve a block of compute time instead of paying on-demand. The blocks are sold in fixed sizes: $36/year for $5/month of credit on shared Machines, up to $3,600/year for $500/month, and $144/year for $20/month on performance Machines, up to $14,400/year for $2,000/month. Three published constraints decide whether that 40% is real on your bill.

• A block is bought per CPU class and per region, and then applies to any number of Machines of that class in that region.
• The credit applies only to CPU and additional RAM charges, never to volumes, snapshots, egress or support.
• The credit does not roll over month to month, and the sizes are fixed, so you rarely match your usage exactly. A $33.00 compute line cannot be reserved at exactly $33.00.

The crossover is still simple.

• Machine runs 24/7 with steady load -> reserve it. Even an imperfectly sized block beats on-demand on a workload you were going to run anyway.
• Machine is spiky or auto-stops -> stay on-demand. Reserving a machine you keep stopping wastes the reservation.

Trap 3: egress is priced by geography

The same byte costs six times more leaving Mumbai than Virginia. Serving 100GB/month of egress from North America is $2.00. Serving the identical 100GB from Africa or India is $12.00. For a globally distributed app, where your traffic lands is a pricing decision, not just a latency one.

Three worked 30-day bills

A. Hobby app, one small machine, always on

The "$5 Hobby plan" the internet quotes does not exist. You pay about $2.74 of metered usage instead, with no free tier to fall back to.

B. Small production API, humans on call

Support is about 44% of this bill. Reserving the machine helps, but not by a clean 40%: performance blocks are sold at $20/month of credit and the credit does not roll over, so two blocks cost $24.00/month amortised and cover the whole $33.00 compute line, taking the bill to about $57.50/mo. One block is $12.00 amortised plus $13.00 of unreserved compute, so about $58.50/mo.

C. The "we scaled to zero" preview environment

Not $0. Delete the volume, not just stop the machine, if a preview is truly dead.

Where Fly.io fits, and where it does not

Fly.io's per-second model rewards workloads that genuinely start and stop, and punishes always-on machines you forgot to reserve. If your app is a steady 24/7 service, price it against usage-billed Railway and instance-billed Render, both of which use different meters that can win at steady load. For container workloads specifically, our northflank-pricing-2026 teardown breaks down a comparable per-resource model.