Competitive positioning: Fly.io versus the alternatives

This page compares Fly.io with the platforms a customer is most likely to name: Render, Railway, Heroku, Vercel, Cloudflare, AWS, Google Cloud Run, DigitalOcean App Platform, and the agent-sandbox vendors E2B, Modal, Daytona, Vercel Sandbox and Cloudflare Sandbox. Competitor facts come only from that vendor’s own public page, fetched 2026-11-04 and linked in the same sentence. Numbers that could not be read from a page are marked [verify]. Nothing here is a claim about anyone’s reliability.

This page is drawn from https://fly.io/docs/about/pricing/, https://fly.io/pricing, https://fly.io/sprites, https://fly.io/docs/sprites/concepts/lifecycle/, https://fly.io/docs/reference/architecture/, https://fly.io/blog/kurt-scott-money-sprites/ and the competitor pages linked below. This guide is unofficial.

The Fly.io angle in one breath

Real VMs that start fast, stop cheaply, and run near users; plus Sprites, persistent Linux computers for agents.

• Firecracker microVMs, not containers on a shared kernel. App code runs in Firecracker microVMs on dedicated servers (https://fly.io/docs/reference/architecture/). Fly’s CEO says the hypervisor engine is the same as Lambda’s (https://fly.io/blog/fuckin-robots/).
• Billed by the second, and stopped Machines cost almost nothing. Started Machines are billed by the second, a stopped Machine bills only its rootfs at $0.15/GB/month, and the smallest always-on Machine (shared-cpu-1x, 256MB) is $2.19 per 30 days (https://fly.io/docs/about/pricing/). There is no free tier for new organizations, only a $1 trial that lasts 7 days or until used up (https://fly.io/docs/about/pricing/).
• Scale to zero with Fly Proxy. With auto_stop_machines and auto_start_machines, Fly Proxy stops idle Machines and starts one when a request arrives. A cold start takes about 2 seconds or more for common apps, and suspend resumes in a few hundred milliseconds (https://fly.io/docs/launch/autostop-autostart/).
• Many regions, one Anycast address. Every app gets a shared IPv4 and Anycast IPv6 at no charge (https://fly.io/docs/about/pricing/), across 17 region codes (https://fly.io/docs/reference/regions/).
• Sprites for agents. “Sprites are persistent, hardware-isolated Linux VMs on Fly.io for AI agents, dev environments and untrusted code, billed by the second and free when idle” (https://fly.io/docs/sprites/).

Be honest about the trade. Fly gives you rawer building blocks than a Heroku-style PaaS: Machines are provisioned up front and started or stopped, not created per request, and autoscaling is metrics-driven automation you own (https://fly.io/docs/reference/aws-to-fly-guide/). Lead with speed, regions and cost shape, not with the line that it is just like Heroku.

Price reference: what each vendor publishes

Entry-level compute only, copied from each vendor’s own page. Sizes differ, so this is a map of billing models, not a ranking.

Agent-sandbox rates as published

Each vendor meters differently (active CPU, allocated CPU, provisioned memory), so treat this as a reading aid, not a bake-off. “Derived” means arithmetic on the published rate.

Render

What it is. A PaaS with web services, workers, cron, Postgres and persistent disks. Plans are workspace tiers plus per-instance compute: Hobby $0, Pro $25/month, Scale $499/month, all ”+ compute” (https://render.com/pricing). Free web services spin down after 15 minutes without inbound traffic and take about one minute to spin back up (https://docs.render.com/free). Render Sandboxes are in beta: “Isolated microVMs for agents, apps, and untrusted code”, with “Sandboxes can run up to 30 days on Pro and above” and 40 GB of disk included per sandbox (https://render.com/pricing).

Where Fly wins

• Regions and Anycast. One app can run Machines in many regions behind one Anycast address (https://fly.io/docs/reference/regions/, https://fly.io/docs/about/pricing/). fly-replay routes requests between regions, Machines or apps (https://fly.io/docs/networking/dynamic-request-routing/).
• Idle cost. A stopped Machine bills only rootfs at $0.15/GB/month (https://fly.io/docs/about/pricing/). Render free web services lose local files when they spin down (https://docs.render.com/free).

Where Render wins

• Git-connected PaaS feel. Render lists “Full-stack PR previews” and “Zero-downtime deploys” as product features (https://render.com/pricing). On Fly, review apps are a GitHub Actions blueprint you set up (https://fly.io/docs/blueprints/review-apps-guide/).
• A free tier. Render has free web services, Postgres and Key Value with limits (https://docs.render.com/free). Fly has no free tier for new organizations (https://fly.io/docs/about/pricing/).
• Managed Postgres ladder. Render lists Postgres plans from $6/month and a free instance (https://render.com/pricing). Fly Managed Postgres starts at $38/month (Basic) (https://fly.io/pricing).

Ask the customer: Do users sit in more than one geography? Do you need a free hobby tier, or a production tier with previews on every PR?

Railway

What it is. Usage-metered hosting: “Railway only charges you for what your app uses”, by the second, with Hobby at $5 minimum usage and Pro at $20 minimum usage, each including that amount as credits (https://railway.com/pricing). It lists Sandboxes as “Isolated Linux VMs for agents, builds, and untrusted code” (https://railway.com/pricing).

Where Fly wins

• Regions are a first-class dial. Fly lists 17 region codes (https://fly.io/docs/reference/regions/). Railway’s plan table lists global regions as “Trial only, unavailable on the free plan” (https://railway.com/pricing).

Where Railway wins

• Elastic metering without choosing a size. “Say goodbye to overprovisioning and optimizing box sizes” (https://railway.com/pricing). On Fly you pick a preset (https://fly.io/pricing).

Ask the customer: Is traffic spiky enough that per-second metering beats picking a Machine size? Do you need multiple regions?

Heroku

What it is. Dynos with fixed monthly prices: Eco $5, Basic $7, Standard-1X $25, Standard-2X $50, Performance-M $250 (https://www.heroku.com/pricing). Eco “Sleeps after 30 minutes of inactivity” (https://www.heroku.com/pricing). Eco requires a $5 flat monthly subscription covering a pool of 1000 hours shared by all Eco dynos (https://devcenter.heroku.com/articles/dyno-types).

Where Fly wins

• Machine sizes and regions. Presets run from shared-cpu-1x to performance-16x, with extra RAM at $6.00 per GB per month up to 128GB (https://fly.io/pricing).

Where Heroku wins

• Buildpack workflow and add-on ecosystem. Heroku markets an “ecosystem of add-ons, integrations, and buildpacks” (https://www.heroku.com/pricing).

Migration note. Fly publishes an Elixir-specific Heroku migration guide (https://fly.io/docs/elixir/getting-started/migrate-from-heroku/); the blog post on it is dated 2020-09-25 (https://fly.io/blog/migrating-heroku-database-apps-to-fly/).

Ask the customer: What does the Procfile look like? Which add-ons are in use, and do they have a Fly equivalent (Managed Postgres, Tigris, Upstash)?

Vercel

What it is. A frontend and functions platform. Hobby $0, Pro $20/month with $20 of included usage, Enterprise custom with a “99.99% SLA” (https://vercel.com/pricing). Functions bill Active CPU (from $0.128 per hour), provisioned memory (from $0.0106 per GB-hr) and invocations (from $0.60 per 1M) (https://vercel.com/pricing).

Where Fly wins

• Long-running and stateful servers. A Machine can run “a 36-hour batch job” (https://fly.io/blog/fuckin-robots/), with volumes for local NVMe storage (https://fly.io/docs/volumes/overview/).
• Any language, any port, private networking. Machines run Docker images as Firecracker microVMs (https://fly.io/docs/reference/architecture/); 6PN and Flycast give private service-to-service networking (https://fly.io/docs/networking/flycast/).

Where Vercel wins

• Frontend developer experience. Automatic CI/CD, a global CDN and a Web Application Firewall are on the Hobby plan list (https://vercel.com/pricing).

Ask the customer: Is this a static or Next.js frontend with a few API routes, or a backend with state, queues or WebSockets? Fly fits the second better; say so plainly.

Cloudflare: Workers, Containers, Durable Objects, Sandboxes

What it is.

• Workers. Free plan: 100,000 requests per day. Standard (Paid, $5 minimum): 10 million requests and 30 million CPU ms included, up to 5 minutes of CPU time per invocation (https://developers.cloudflare.com/workers/platform/pricing/). “There are no additional charges for data transfer (egress) or throughput (bandwidth)” (https://developers.cloudflare.com/workers/platform/pricing/).
• Containers. Billed per 10 ms while active. Instance types run from lite (1/16 vCPU, 256 MiB) to standard-4 (4 vCPU, 12 GiB, 20 GB disk). Requests reach a container through a Worker and a Durable Object, and you are billed for both (https://developers.cloudflare.com/containers/pricing/).
• Durable Objects. Paid plan: 1 million requests per month then $0.15 per million, and 400,000 GB-s per month then $12.50 per million GB-s of duration (https://developers.cloudflare.com/durable-objects/platform/pricing/).
• Sandboxes. “Each instance is a microVM with its own kernel and network, so no other workload shares it”, available on the Workers Paid plan (https://developers.cloudflare.com/sandbox/).

Where Fly wins

• Your existing app runs as is. Fly says it moved from a JavaScript edge runtime to containers because customers “just wanted their native code to work” (https://fly.io/blog/wrong-about-gpu/). Cloudflare Containers still sit behind a Worker and a Durable Object (https://developers.cloudflare.com/containers/pricing/).
• Persistent disk and long-lived processes without the Worker front door: volumes (https://fly.io/docs/volumes/overview/) and Machines that “stick around as long as you want” (https://fly.io/blog/fuckin-robots/).

Where Cloudflare wins

• Egress. Workers have no data transfer charge (https://developers.cloudflare.com/workers/platform/pricing/); Fly bills outbound data at $0.02/GB in North America and Europe (https://fly.io/docs/about/pricing/).
• A free tier. Workers Free exists (https://developers.cloudflare.com/workers/platform/pricing/); Fly has none for new organizations (https://fly.io/docs/about/pricing/).

Ask the customer: Is the code already written for the Workers runtime? Is a per-request, CPU-time model the right shape for it, or does it need a normal Linux process?

AWS: Fargate, Lambda, App Runner, Lightsail

What it is.

• Fargate. Fargate Spot runs interrupt-tolerant Amazon ECS tasks on spare capacity at up to a 70% discount off the regular Fargate price, per the page. Regular per-vCPU and per-GB rates could not be read from the page: [verify] (https://aws.amazon.com/fargate/pricing/).
• Lambda. The page now lists three shapes: Event Functions (per request and duration), MicroVMs (per instance-second) and Web Functions (per request and active CPU time). Rates [verify] (https://aws.amazon.com/lambda/pricing/).
• App Runner. $0.007 per GB-hour for provisioned memory and $0.064 per vCPU-hour plus $0.007 per GB-hour when active, in the four regions the page names (https://aws.amazon.com/apprunner/pricing/).
• Lightsail. Fixed bundles from $5/month for 0.5 GB, 2 vCPUs, 20 GB SSD and 1 TB transfer (https://aws.amazon.com/lightsail/pricing/).

Where Fly wins

• The trade, in Fly’s own words: “You get rawer access to your infra and better latency for your users, but you might trade some convenience” (https://fly.io/docs/reference/aws-to-fly-guide/).

Where AWS wins

• Breadth. Fly’s own blog lists S3, SQS, ECS, Lambda, EBS, ALBs, CloudWatch and more around AWS compute (https://fly.io/blog/carving-the-scheduler-out-of-our-orchestrator/).
• Native alerting and managed integrations. Fly’s docs say there is “no native alerting yet” and point to Prometheus, Alertmanager or Grafana (https://fly.io/docs/reference/aws-to-fly-guide/).

Ask the customer: How many AWS services does the app touch? If it is RDS, S3 and little else, Managed Postgres and Tigris are the Fly answers (https://fly.io/docs/reference/aws-to-fly-guide/).

Google Cloud Run

What it is. “Cloud Run charges you only for the resources you use, rounded up to the nearest 100 millisecond.” Free tier for instance-based services: first 240,000 vCPU-seconds and 450,000 GiB-seconds free per month; for request-based billing it also includes 2 million requests (https://cloud.google.com/run/pricing). Services “scale to zero when idle”, with minimum and maximum instance limits you can set (https://cloud.google.com/run/docs/about-instance-autoscaling). Per-unit rates are region-dependent and not readable from the fetched page: [verify].

Where Fly wins

• Volumes. Local NVMe volumes attach to Machines (https://fly.io/docs/volumes/overview/).

Where Cloud Run wins

• Configurable request concurrency per instance. Fly’s own blog notes “Google’s Cloud Run product supports configurable per-instance concurrency” (https://fly.io/blog/the-serverless-server/). Fly has soft_limit and hard_limit concurrency settings (https://fly.io/docs/blueprints/setting-concurrency-limits/).
• A free tier and committed use discounts across GKE, Compute Engine and Cloud Run (https://cloud.google.com/run/pricing).

Ask the customer: Are you already inside Google Cloud IAM, VPC and Artifact Registry? That is the strongest reason to stay.

DigitalOcean App Platform

What it is. Container components with fixed monthly prices: shared 1 vCPU 512 MiB $5.00, shared 1 vCPU 1 GiB $12.00, dedicated 1 vCPU 512 MiB $29.00 (https://docs.digitalocean.com/products/app-platform/details/pricing/). “Apps are billed by the second, with a one-minute minimum” (https://docs.digitalocean.com/products/app-platform/details/pricing/). A free tier covers 3 static-site apps (https://www.digitalocean.com/pricing/app-platform).

Where Fly wins

• Idle and regions. Stopped Machines stop CPU and RAM billing (https://fly.io/docs/launch/autostop-autostart/), and apps span many regions (https://fly.io/docs/reference/regions/).

Where DigitalOcean wins

• Bundled bandwidth. The $5 shared container includes 50 GiB per month of transfer (https://docs.digitalocean.com/products/app-platform/details/pricing/).

Ask the customer: Is the workload one small app plus a database, where predictable monthly tiers matter more than region reach?

Agent sandboxes: Sprites versus E2B, Modal, Daytona, Vercel Sandbox, Cloudflare Sandbox

Fly’s position. Fly’s CEO argues that agents want computers rather than sandboxes, in a post that also introduces a new iteration of Sprites and a company focus on them (https://fly.io/blog/kurt-scott-money-sprites/). A Sprite has an ext4 disk, 8 vCPUs, 100 GB of storage, goes warm after about 30 seconds idle (resume in 100-500 ms with processes preserved) and cold later (wake in 1-2 s, processes fresh); open TCP connections do not survive either pause (https://fly.io/docs/sprites/concepts/lifecycle/). Sprites do not autoscale storage yet (https://fly.io/docs/sprites/concepts/lifecycle/).

Where Fly wins

• Idle is nearly free and the disk is real. Compute stops billing when idle; cold storage is $0.000027 per GB-hour, about $0.02 per GB-month (https://fly.io/pricing).
• Credentials stay out of the box. Sprite Connectors let calls reach GitHub, OpenRouter, Slack or any HTTP API without the credential entering the Sprite (https://fly.io/sprites).

Where the others win

• E2B: pause keeps memory state too, and the pricing page lists SOC 2 and HIPAA (https://e2b.dev/docs, https://e2b.dev/pricing). Fly Sprites keep disk, not memory, across a cold wake (https://fly.io/docs/sprites/concepts/lifecycle/).
• Modal: GPU pricing from T4 to B300 is listed on the same pricing page (https://modal.com/pricing). Fly says it will not do “something bigger” with GPU Machines, and a v2 is unlikely soon (https://fly.io/blog/wrong-about-gpu/).
• Daytona: lists GPU, Windows and macOS sandboxes (https://www.daytona.io/docs/en/sandboxes.md).
• Vercel Sandbox and Cloudflare Sandbox: sit next to the customer’s Vercel or Workers app (https://vercel.com/docs/vercel-sandbox/pricing, https://developers.cloudflare.com/sandbox/).
• Render and Railway sandboxes: Render documents sandboxes that “can run up to 30 days on Pro and above” and Railway lists 100 running sandboxes per environment on Pro (https://render.com/pricing, https://railway.com/pricing).

Ask the customer: Does the agent need state across idle gaps? Does it need GPUs, Windows, or memory-level resume? Those decide it.

Honest limits to bring up first

Objection: “Will Fly.io be around?”

Fly’s own CEO addressed this in public. A reviewer named Fly as the provider he was least confident would be around by the end of the year; Fly replied that it was in a run of strong quarters, had raised more money and named Scott Johnston CEO (https://fly.io/blog/kurt-scott-money-sprites/). Do not argue the point. Say: “Here is Fly’s own post on it.” Then offer the contract-level answer: Enterprise support starts at $2,500/month (https://fly.io/docs/about/pricing/), and ask whether an SLA or exit plan is the real concern.