Micro-PaaS Reality Check 2026 — Fly.io, Heroku, and Render

Mon, 13 Apr 2026 00:00:00 +0900

Overview

For the small CPU-bound side of an app — the API server, the worker queue, the Postgres database — is a micro-PaaS still cheaper than rolling EC2? In 2026, the answer is “almost always, until you cross $200/month.” This post compares Fly.io, Heroku, and Render, and gives a decision framework for when to walk away from PaaS entirely.

The Three Platforms at a Glance

graph TD
 A[App needs hosting] --> B{Need global edge?}
 B -->|Yes| C[Fly.io <br/> Firecracker microVMs]
 B -->|No| D{Need managed Postgres <br/> + add-on ecosystem?}
 D -->|Yes| E[Heroku <br/> classic PaaS]
 D -->|No, want simplicity| F[Render <br/> modern Heroku alt]
 A --> G[Heavy traffic <br/> or specialized infra?]
 G -->|Yes| H[EC2 + Terraform]

Fly.io

Fly runs your Docker images on Firecracker microVMs across 35+ global regions. Pricing is roughly $0.0000022/sec per shared-cpu-1x VM (~$1.94/mo if always-on for a 256MB instance), and you can scale to zero on certain plans. The killer feature is fly.toml plus flyctl deploy — git-push-style deploys without CI/CD pipelines.

# fly.toml
app = "my-api"
primary_region = "nrt"

[http_service]
 internal_port = 8080
 force_https = true
 auto_stop_machines = true
 auto_start_machines = true
 min_machines_running = 0

Postgres is unmanaged-by-Fly (you run their image yourself); for a managed alternative they now point users to Supabase or Neon.

Best for: geographically distributed apps, anyone who wants Firecracker isolation, projects where TCP/UDP (not just HTTP) matters.

Heroku

The granddaddy of PaaS, now under Salesforce. The 2026 platform has two foundations:

Cedar — the classic dyno (LXC-based, broad add-on compatibility)
Fir — Kubernetes-powered, more observability and finer control

Tier	Price	Use case
Eco dyno	`$5/mo`	Hobby / staging
Basic	`$7/mo`	Small production apps
Standard-1X	`$25/mo`	Real production
Heroku Postgres essentials	`$5/mo`	10K rows

Add-ons go through the Elements Marketplace where 1 enterprise credit = $1.

The new bet is Heroku Managed Inference and Agents — a curated set of LLMs (text-to-text, embedding, image generation) plus MCP server hosting on pay-as-you-go dynos. This is Heroku trying to be the “easy AI app deployment” platform. Whether it competes with Vercel AI SDK + Modal-style stacks is an open question, but Heroku has the deployment ergonomics to make it credible.

Best for: apps that need a real managed Postgres, teams with low ops budget, anyone who wants git push heroku main with zero config.

Render

The Heroku alternative everyone migrated to during the free-tier shutdown of 2022. Render advertises Heroku migration credits up to $10K. Pricing is competitive:

Service	Price
Static sites	Free tier
Web services	From `$7/mo`
Managed Postgres	From `$7/mo`
Background workers	From `$7/mo`
Cron jobs	Free

Native support for cron jobs, background workers, and preview environments. Render Workflows is their newer orchestration layer for multi-service deploys.

Best for: former Heroku users, teams who want preview environments out of the box, projects that need Docker support without the Fly.io geo-distribution complexity.

Side-by-Side

Capability	Fly.io	Heroku	Render
Global edge	✅ 35+ regions	❌ US/EU only	❌ US/EU only
Managed Postgres	❌ (Supabase/Neon)	✅ first-party	✅ first-party
Scale-to-zero	✅	❌ (Eco can sleep)	❌
Docker native	✅	✅ (Fir)	✅
Preview envs	⚠️ via flyctl	✅ Pipelines	✅ Workflows
Cron / workers	⚠️ separate machines	✅	✅
AI/LLM hosting	❌	✅ Managed Inference	❌
Cheapest always-on tier	`~$2/mo`	`$5/mo`	`$7/mo`

Decision Framework

flowchart LR
 A[Monthly bill projection?] -->|<$25| B[Take any PaaS <br/> happily]
 A -->|$25-$200| C[Pick by feature fit]
 A -->|>$200| D[Consider EC2 + Terraform <br/> if team has ops bandwidth]
 C --> E[Postgres-heavy: Heroku/Render]
 C --> F[Global users: Fly.io]
 C --> G[Heroku migration: Render]

A useful heuristic: if your app fits in $25/mo, take the managed PaaS happily. The hour you save not configuring Terraform and Nginx is worth more than the platform markup. Once you cross $200/mo of PaaS billing, EC2 + a thin Terraform module starts being the cheaper path — but only if someone on the team enjoys ops.

What About Vercel and Railway?

Worth naming them as adjacent options:

Vercel dominates the Next.js / frontend deployment niche. For an SSR React app, it’s the default. For a Python API or Go service, you’re better off elsewhere.
Railway is the slickest DX of the bunch, but pricing has shifted upward post-pivot; it’s no longer the “obvious cheap” option it was in 2023.

Insights

The cloud-cost narrative of 2024-2025 (“everyone’s moving back to bare metal!”) is mostly noise for small teams. At small scale, the markup of managed platforms is lower than the engineering cost of replacing them. Fly.io continues to be the developer-experience benchmark, Heroku is genuinely back from the dead with Fir + Managed Inference, and Render is the boring-correct choice for most CRUD apps. The right framing isn’t “PaaS vs EC2” — it’s “PaaS until your bill or your scale forces a migration.” For most small apps that day never comes.

Paas on ICE-ICE-BEAR-BLOG