Free compute can cover a small API, a database, an occasional GPU job or even a Linux server. The awkward part is that all four can be advertised as “free hosting” while giving you very different things. Start with what needs to run, then check what happens when the allowance runs out.
This page is maintained in place: it replaces the June guide and the July to October monthly editions, and the changelog at the bottom records what moved each month. The provider documentation linked below was checked on 12 November 2026. These are documentation checks, not a claim to have opened and tested every account. Regional capacity, account eligibility and provider terms still matter.
Which kind of free compute fits the job?
Compare the job, the recurring allowance and what happens at the limit. On a narrow screen, scroll across the table for the limit and billing conditions.
| Job and service | Allowance and reset | When it runs out | Card or billing | Main constraint |
|---|---|---|---|---|
| Linux VM: Oracle Always Free | Up to two AMD micro VMs; Ampere A1: 1,500 OCPU-hours and 9,000 GB-hours/month; 200 GB combined boot/block storage; 10 TB outbound/month. | Free Tier accounts are not charged; after a paid upgrade, use above the Always Free limits is billed. | Card usually needed for verification; paid upgrade is optional. | Regional capacity is not guaranteed; idle VMs can be reclaimed. |
| Small Linux VM: Google Cloud e2-micro | One e2-micro instance-month in us-west1, us-central1 or us-east1; 30 GB-months standard disk; 1 GB North America egress. Unused amounts do not roll over. | Billed to a paid billing account, or taken from trial credit. | Cloud Billing account required; separate $300 credit for 90 days. | Three US regions only; a tiny machine. |
| Agent or dev VM: Freestyle Free | 200 vCPU-hours, 400 GiB-hours memory and 50 GB transfer, reset monthly; up to 10 concurrent VMs. | VMs pause; you resume them after the reset. | $1 card verification, refunded; Free never bills usage. | Billed on allocated resources, busy or idle. |
| Small API or webhook: Cloudflare Workers Free | 100,000 requests/day, reset at midnight UTC; 10 ms CPU/request, 128 MB memory. | Requests return an error or bypass the Worker until the reset. | Free plan; paid usage requires an upgrade. | Short CPU budget. |
| Web service: Render Free | 750 instance hours per workspace per calendar month. | Free web services are suspended until the next month. | Bandwidth overage is billed with a payment method; without one, Free services are suspended. | Spins down after 15 idle minutes; about a minute to wake; no persistent disk. |
| Intermittent Postgres: Neon Free | 100 CU-hours and 5 GB transfer per project each billing period; 1 GB Postgres/project, 20 GB account total. | Compute is suspended until the next period; writes fail at the storage cap; data is kept. | No card; no Free-plan overages. | Sleeps after five idle minutes. |
| Managed database: Aiven Free | Free PostgreSQL, Kafka, OpenSearch or Valkey service with no time limit; resource limits vary by service. | A fixed service size, not a metered allowance. | No card; no fixed trial expiry. | Inactive services can be powered down; cloud, region or configuration can change. |
| Short agent task: Vercel Sandbox Hobby | 5 active CPU-hours, 420 GB-hours memory, 5,000 creations and 20 GB transfer/month. | New sandbox creation pauses; no Hobby overage charges. | Hobby plan; personal, non-commercial use. | 45 minutes/session, ten concurrent. |
| Burst CPU or GPU jobs: Modal Starter | $30 compute credit/month; hours depend on hardware and workload. | Usage beyond the credit is billed unless a spend limit stops it. | Payment method required. | Token-billed Shared Endpoints are excluded from the credit. |
| Hosted AI inference: Cloudflare Workers AI | 10,000 Neurons/day, reset at midnight UTC; inference volume depends on the model. | Free: further requests fail. Paid: $0.011 per 1,000 Neurons. | Several large models require paid billing or prepaid credits. | This is an API, not your own GPU. |
A free plan that stops at a limit is easier to budget than a free allowance inside a billable account. A trial credit has an expiry or a finite balance. Shared capacity may be free and still unavailable when you want it. Keep those distinctions attached to the recommendation.
Free VPS hosting: Oracle Cloud Always Free
Oracle's Always Free resource catalogue includes up to two AMD micro VMs, each with 1 GB of memory and up to 50 Mbps of internet bandwidth, and an Ampere A1 allowance of 1,500 OCPU-hours and 9,000 GB-hours each month. Oracle describes the latter as equivalent to 2 OCPUs and 12 GB RAM, usable as one instance or two smaller ones. The tenancy also has 200 GB of combined boot and block storage and 10 TB of outbound transfer a month.
The catch is capacity in your home region. An advertised allocation is not a reservation, and an “out of host capacity” error means the free shapes are temporarily unavailable. Oracle can also reclaim an instance that is idle for seven days, which it defines as 95th-percentile CPU, network and (on A1) memory use all below 20%. Those constraints make it useful for experiments you can recover, but a poor basis for promising somebody an uninterrupted service for nothing.
If you already own a spare machine, my experience of making a home server useful covers the other route, including why self-hosting everything was not the answer. Electricity and maintenance still count.
Other free Linux VMs, and the temporary ones
Google Cloud's free tier includes one non-preemptible e2-micro VM a month in Oregon, Iowa or South Carolina, with 30 GB-months of standard persistent disk and 1 GB of outbound transfer from North America. It is small, but it does not expire. The separate $300 Welcome credit lasts 90 days, so treat it as a trial rather than a hosting plan.
Freestyle's Free plan is aimed at agents and development environments. It now publishes monthly allowances: 200 vCPU-hours, 400 GiB-hours of memory, 60,000 GiB-hours of storage and 50 GB of transfer, with up to ten concurrent VMs, ten saved VMs and ten snapshots. Each VM can have up to 4 vCPUs, 8 GiB of memory and a 32 GB disk. The allowances are hard caps: VMs pause when compute runs out, and you can resume them after the reset. Adding a card triggers a refunded $1 verification charge.
Railway's anonymous SSH boxes are the quickest route to a shell: ssh railway.new gives you 2 vCPU and 2 GB RAM with coding agents preinstalled and a preview URL. Unclaimed boxes and their files are deleted after 24 hours, and there is a limit of three per IP address per day. The ordinary account trial is a one-time $5 credit for 30 days, then the Free plan gives $1 of credit a month. Volumes from trial accounts are deleted 30 days after the credit expires.
Daytona advertises $200 of free compute with no credit card, plus 5 GiB of free storage, billed per second after that. Its pricing and billing pages do not say whether that credit recurs, when it expires or who qualifies, and free credit cannot be spent on GPU sandboxes. Until Daytona documents a monthly allowance, treat it as a finite credit: it suits a burst of sandbox work rather than a long-running server.
Hyperscaler free tiers and trial credits
The big clouds mostly offer credit with a clock on it, plus a list of small always-free services.
- AWS: new accounts get up to $200 of credit, $100 at signup and up to $100 more for using key services. A Free plan account closes after six months or when the credit runs out, whichever comes first, unless you upgrade. More than 30 services stay free within monthly limits.
- Google Cloud: the e2-micro VM above, Cloud Run's request-based free tier (two million requests, 360,000 GB-seconds of memory and 180,000 vCPU-seconds a month), and the 90-day $300 credit.
- Azure: 12 months of free monthly amounts for new customers, 65+ always-free services, and a separate $100 student credit with no card. You need to move to pay-as-you-go within 30 days to keep the 12-month services.
- IBM Cloud: Lite plans cover 40+ always-free products and never expire, but IBM does not list an always-free virtual server. The $200 credit lasts 30 days, and payment details are required up front.
None of these is a free VPS in the Oracle sense. They are a good way to try a managed service, and a poor way to host something you intend to leave running.
Small APIs: Cloudflare Workers and Cloud Run
Workers Free includes 100,000 requests per day, 10 milliseconds of CPU time per HTTP request, 128 MB memory and 50 subrequests per request. Worker execution stops at the daily request limit until the quota resets at midnight UTC. Fail-closed routes return errors; fail-open routes bypass the Worker.
That is a practical shape for a small API, webhook or bit of application glue. The CPU limit is the important qualification: a high request count does not turn it into a general-purpose VM. Storage and other Cloudflare products have their own allowances. The Cloudflare Workers free tier guide works through a complete example, and free cron jobs covers scheduled tasks.
Cloud Run is the container alternative: two million requests a month with a monthly pool of vCPU and memory seconds. It suits a patient batch job or an API that is fine with cold starts, inside a billing account you should cap.
Render's free web services are closer to an ordinary app host. Each workspace gets 750 instance hours a month. A service spins down after 15 minutes without traffic and takes about a minute to wake, local file changes are lost when it spins down, restarts or redeploys, and the free Postgres database expires after 30 days. When the workspace uses all 750 hours, Render suspends its Free web services until the next month. Koyeb no longer lists a free compute instance; its pricing page shows only a small free Postgres allowance.
Free databases: Neon, Aiven and the rest
Neon's Free plan allows up to 100 projects. Each gets 100 CU-hours of compute per month, autoscaling up to 2 CU, 1 GB of Postgres storage and 5 GB public network transfer shared with the project's Object Storage and Functions. Postgres storage also has a 20 GB account-wide total. Compute scales to zero after five minutes of inactivity. Exhausting compute or transfer suspends it until the next billing period or an upgrade; reaching the storage cap makes writes fail. None of these limits deletes data, and the Free plan does not bill overages.
Free also includes 5 GB of Object Storage per project and monthly Functions allowances of 10 active Capacity-Hours, 400 waiting Capacity-Hours and one million invocations per project. Functions concurrency has separate account-wide limits, and since 23 October a function's environment variables are limited to 32 KiB in total.
That makes sense for an intermittent Postgres application with a small dataset. It is less comfortable for something that must answer continuously, because the compute budget is finite and sleeping is part of the plan.
Aiven's free services cover PostgreSQL, Kafka, OpenSearch and Valkey without a credit card or a fixed trial period. Aiven has retired its free MySQL tier, so new MySQL services start on a paid plan. Aiven can power down a free service with no initial or continuing activity, and can change its cloud provider, region or configuration. Choose the actual database your project needs, then check its individual service limits.
For a specific database, the deeper guides cover signup, limits and recovery: Supabase, Turso, Prisma Postgres and free Redis hosting.
Agent sandboxes: Vercel Sandbox
Vercel Sandbox on Hobby includes five active CPU-hours, 420 GB-hours of provisioned memory, 5,000 creations and 20 GB of data transfer per month. Data a sandbox downloads, such as packages and repositories, is free; outbound data and traffic through exposed ports counts. A single session can run for up to 45 minutes, with ten sandboxes running concurrently. Exceeding the included quota pauses new sandbox creation rather than charging Hobby overages.
Vercel Drives preserve a mounted workspace across runs. Hobby includes 15 GB of Drive storage and monthly allowances of 30 GB of reads and 30 GB of writes. Drives do not increase the compute pool. Hobby is for personal, non-commercial use.
CPU-hours and memory-hours are different budgets. Waiting on an API uses little CPU while the allocated memory continues to count. Check both before assuming that a quiet agent can stay running all month.
Sandbox sessions or an always-on VPS?
A sandbox and a free VPS can both give you a shell, but they behave differently when nothing is happening. A VPS keeps its processes running until you stop it or the provider reclaims it. A sandbox is built to stop, pause or time out, and its free compute is a metered allowance or a credit, so leaving one running uses it up.
- Vercel Sandbox has no idle stop: a session runs until its timeout, at most 45 minutes on Hobby, and its memory counts while it waits. Sandboxes are persistent by default. Stopping one snapshots the filesystem, and the next session boots from that snapshot without the previous processes. Snapshots expire 30 days after their last use unless you change that, and Hobby includes 15 GB of snapshot storage.
- Daytona stops a sandbox after 15 minutes without activity unless you change or disable the interval. Previews, SSH sessions and Toolbox API calls count as activity; a background script on its own does not. A stopped sandbox keeps its filesystem but not its memory and is billed for disk only. Container sandboxes are archived after seven days stopped, and archived ones are not billed.
- Freestyle bills a running VM on what it has allocated, busy or not. Its idle timeout is something you configure. Pausing saves memory, so processes resume where they were; stopping keeps the disk and discards memory. Even a one-vCPU VM would use the 200 vCPU-hours in a little over eight days.
- Railway's unclaimed SSH boxes are deleted with their files after 24 hours.
Use a sandbox for work that starts, does something and stops: an agent task, a test run or a build. Keep the files that matter in a snapshot or drive, and restart services when it resumes. Use an always-free VPS such as Oracle's when a process has to keep listening, and accept its capacity and idle-reclaim rules. Neither Daytona nor Freestyle documents a free plan meant to run a server continuously: Freestyle's monthly compute runs out after days of continuous use, and Daytona's pages do not say whether its credit recurs.
GPU jobs and hosted AI APIs
Modal's Starter plan includes $30 of compute credit each month for Functions and Sandboxes, plus 1 TiB of Volumes storage and 1 TiB of egress. Its CPU, memory and GPU prices determine how much work that buys: at the listed per-second rates, $30 is roughly 50 hours of a T4 or under eight hours of an H100, before CPU and memory. There is no honest single number of “free GPU hours” without specifying the job. The credit excludes token-billed Shared Endpoints, which charge from the first request. Modal requires a payment method and can charge beyond the credit. Set a Workspace spend limit before leaving jobs running.
Cloudflare Workers AI includes 10,000 Neurons per day, resetting at midnight UTC. Neurons are Cloudflare's usage unit, and the amount of inference that buys depends on the model. On Workers Free, further operations fail after the allowance is exhausted; Workers Paid charges $0.011 per 1,000 Neurons beyond it. Several of the largest models, including the current Kimi, GLM and DeepSeek V4 variants, require a paid billing method or prepaid AI Gateway credits. The Workers AI free tier guide has the details.
For hosted model APIs without a GPU of your own, see Groq's free tier limits, OpenRouter's free models and free embedding APIs.
Before building around an allowance
Check whether the plan permits your intended use, whether a card is required, and whether hitting a limit stops work or starts a bill. Then check the less exciting parts: disk, public IP addresses, outbound traffic, idle shutdown and data retention. Keep a backup or an export route for anything you would mind losing.
This page stays at the same address as the shortlist changes, so it is the useful one to bookmark or link to.
Changelog
12 November 2026. The comparison table now shows each allowance's reset, what happens when it runs out, and the card or billing condition in separate columns. Added a section on sandbox sessions versus an always-on VPS. Aiven has retired its free MySQL tier. When Render's 750 hours run out, Free web services are suspended until the next month whatever the payment method; the payment-method rule applies to bandwidth overage. Freestyle VMs paused at the cap must be resumed after the reset rather than resuming on their own. Daytona's pages do not say whether its $200 credit recurs or expires, so the guide no longer calls it a non-recurring credit. The Vercel Sandbox entry now notes that sandboxes keep their filesystem between sessions by default. Oracle, Google Cloud, Workers, Workers AI, Neon, Modal and Railway allowances were unchanged.
November 2026. Freestyle now documents monthly Free allowances (200 vCPU-hours, 400 GiB-hours of memory, 50 GB transfer) and pauses VMs at the cap. Neon's plan page lists up to 100 Free projects and autoscaling to 2 CU, and Functions environment variables are capped at 32 KiB from 23 October. Workers AI names the large models excluded from the free allocation. Added Google Cloud's e2-micro, Render, Railway's trial terms, Daytona and the hyperscaler credit offers. Oracle, Workers, Vercel Sandbox and Modal allowances were unchanged. This page also absorbed the June guide and the monthly editions.
October 2026. Railway added anonymous SSH VMs, and Neon published Free backend quotas after general availability. Vercel Sandbox Drives began preserving workspaces between runs while Hobby compute and session limits stayed the same. Modal's Shared Endpoints, Lightning's starter credits and Freestyle's persistence needed separate qualifications.
September 2026. Deno Deploy reduced several Free allowances, and Cloudflare D1 began stopping queries at its daily limits. Railway, Cloudflare and Axiom offered temporary infrastructure before signup. Vercel Sandbox published clear recurring quotas; Freestyle, Browserless, Firebase Authentication and Axiom entries were corrected.
August 2026. GitHub Models was retired, SageMaker Studio Lab closed to new users, and Hugging Face's general free CPU Spaces claim no longer held. Paperspace documented free M4000 notebook access. Google Agent Engine, Cloudflare Browser Run, Browserbase and Browserless added recurring agent and browser allowances.
July 2026. Oracle's own pages disagreed on whether Ampere includes 2 OCPUs and 12 GB RAM or twice that; the resource catalogue now says 2 and 12. Upstash, Vercel, Freestyle and Cloudflare offered recurring allowances for agent sandboxes and browser automation.
June 2026. First guide. The rule from then still holds: claim durable free primitives such as Oracle Always Free and Cloudflare before relying on short-lived trial credits, and date every free-tier decision.
More free resources for a side project
Once you have somewhere to run it, free APIs for side projects covers useful services to build with, free datasets and free maps cover data, and free assets for side projects covers icons, art, fonts and sound. If the project is a static site, GitHub Pages limits is worth a look. Check each resource's usage terms alongside the hosting allowance.