Fly.io Pricing in 2026: Plans, Cost & Free Trial

Yes — no credit card required for the free monthly allowance. Fly.io provides 3 shared-CPU VMs (256MB RAM each), 3GB of volume storage, and 160GB of outbound bandwidth every month at no cost. This is enough to deploy a test AI application, explore the CLI and deployment workflow, and evaluate Fly.io's architecture. For AI services requiring larger machines, GPU access, or more volume storage, a payment method is required. Creating an account and deploying within the free allowance takes under 15 minutes with the Fly CLI.

Fly.io charges for the machines running in each region — no per-region infrastructure fee. Adding a second region means adding machine instances in that region, billed at the same per-second rate as your primary region. A performance-2x machine running in 3 regions (US-East, EU-London, Asia-Singapore) costs roughly 3x a single machine — about $92/month total. There are no load balancer fees, VPC charges, or regional networking overhead charges. Inbound data is free; outbound bandwidth from all regions combined counts toward your monthly allowance. This makes multi-region AI deployment significantly cheaper on Fly.io than equivalent AWS or GCP multi-region setups.

Fly.io does not charge for deploys or builds. You pay only for running machine time, volume storage, and outbound bandwidth. When you deploy a new version of your AI application, Fly.io briefly runs both the old and new machines simultaneously during the cutover — you're charged for that overlap, but it typically lasts under 60 seconds and costs fractions of a cent. Docker image builds happen on Fly.io's build infrastructure at no charge. For AI applications with frequent deployments during active development, this zero-cost deployment model keeps iteration costs minimal.

Fly.io GPU Machines and RunPod serve different use cases. RunPod's serverless GPU pricing charges per inference request with a per-second billing model — better for intermittent or batch GPU workloads where you don't want to pay for idle GPU time. Fly.io GPU Machines are always-on (you pay for the machine running, not per request) — better for real-time inference APIs that need consistent availability without cold starts. For always-available LLM inference where users make requests continuously, Fly.io's model provides consistent latency. For batch processing or infrequent inference where per-request billing is more economical, RunPod's serverless is typically cheaper. Contact Fly.io for current GPU Machine pricing.

Volume storage requirements depend entirely on your model. At $0.15/GB/month, storage is relatively inexpensive. Rough guidelines: a 7B parameter model with 4-bit quantization (Q4_K_M) requires approximately 4-5GB. A 13B model with Q4 quantization needs 7-8GB. A full-quality Llama 3 70B model requires 40+GB for 4-bit quantized weights. Text embedding models (all-MiniLM, nomic-embed) are much smaller at 100-500MB. For a service running a single 7B model, a 10GB volume ($1.50/month) provides adequate space with room for the model, any fine-tuning adapters, and storage for cached data.

Standard Fly.io machines are not preempted — they run continuously until you stop or destroy them. This is an important distinction from AWS spot instances or RunPod Community Cloud pods, which can be reclaimed with short notice during high demand. Fly.io's non-preemptible machines are more expensive than spot alternatives, but AI inference requests in progress are never interrupted by infrastructure events. For production AI APIs where mid-inference interruption would cause user-facing errors, Fly.io's consistent availability is worth the premium over preemptible compute options.

Machine size depends on the model and quantization level. For 7B parameter models with 4-bit quantization, a performance-2x machine (4GB RAM) provides adequate memory with some headroom. For 13B models with Q4 quantization, plan for 8GB+ RAM (performance-4x or larger). For 70B models, you'll need GPU Machines — CPU inference at that scale is impractically slow for real-time use. Generally, allocate RAM equal to the model's quantized file size plus 1-2GB overhead for the inference server and OS. Monitor actual RAM usage after deployment and resize if you're approaching the ceiling.

Yes — Fly.io has a startup program providing compute credits to qualifying early-stage companies. AI startups from recognized accelerators (Y Combinator, Techstars, and others) typically qualify. Credits offset compute costs during early product development, reducing the runway impact of infrastructure spending. The application process is through Fly.io's website. In addition to financial credits, accepted startups often receive access to the Fly.io team for technical guidance on architecture and optimization. For pre-revenue AI startups where every dollar of runway matters, applying for Fly.io credits before committing to monthly compute spend is worth the effort.

Outbound bandwidth beyond the 160GB monthly free allowance is billed per GB. For AI response streaming, each LLM response consumes bandwidth proportional to its length. A 1,000-token text response is approximately 4KB; a 4,000-token response is about 16KB. A moderately active AI chat API handling 50,000 requests per month at an average 2,000 tokens per response generates roughly 400MB of outbound data — well within the free bandwidth allowance. High-volume AI APIs handling millions of requests monthly with verbose responses (code generation, long-form content) should calculate expected bandwidth and include it in cost estimates. Current bandwidth rates are at fly.io/pricing.