Free compute in October 2026: new backends and temporary VMs

October’s useful additions are a Linux development machine you can open without signing up, a free backend with published quotas, and persistent storage for agent sandboxes. Railway, Neon and Vercel all made dated announcements in September. There are also corrections to make: Lightning’s current GPU credit offer is a starter balance, Freestyle does include ordinary snapshots, and Modal’s monthly compute credit does not cover its Shared Endpoints API.

This is the 1 October 2026 check. The maintained free-compute guide has the shortlist to bookmark; the September edition is the previous snapshot. This edition concentrates on what changed and what needs correcting, rather than reprinting every provider.

The figures were checked against official documentation and pricing pages. No accounts or workloads were created for this check. A dated announcement establishes when something launched; a changed pricing page without that history only establishes what the provider says today.

A Linux VM before the signup form

Railway launched anonymous SSH VMs on 25 September. The entry point is short enough to be useful:

ssh railway.new

That opens a Linux VM with 2 vCPU and 2 GB RAM, coding agents preinstalled, and a private preview URL. No account or payment card is required. September’s anonymous offer covered sites and databases; this gives you a shell and a machine to build on.

You get 60 minutes to work, followed by 24 hours to claim the VM. Claiming moves the files, VM and URL into an account and makes the preview shareable. Unclaimed machines and files are deleted.

The current trial documentation limits anonymous access to three boxes per IP address per day, with regional capacity restrictions. A claimed workload moves into normal account credits and billing. Railway’s separate signup trial remains $5 for 30 days, followed by $1 of recurring monthly Free credit.

Use the anonymous machine for a disposable development session. Save or claim work before the clock runs out.

Neon’s free backend has actual numbers now

Neon’s backend became generally available on 18 September. The September article described Functions and Object Storage as free during beta in Ohio, with unpublished guardrails. That description is now out of date: the backend has published Free allowances, and those products are available in Ohio, Northern Virginia, Frankfurt and Singapore.

The current Free pricing includes:

• Postgres: up to 100 projects, each with 100 CU-hours per month and 0.5 GB database storage.
• Object Storage: 5 GB per project.
• Functions: 10 active Capacity-Hours, 400 waiting Capacity-Hours and one million invocations per month.
• Auth: 60,000 monthly active users.

The easy mistake is multiplying all those allowances by the project count. Functions limits apply across the account as well as independently per project. The same plan documentation says a project’s 5 GB network allowance is shared by Postgres, Object Storage and Functions.

Active and waiting Capacity-Hours are separate meters for time doing CPU work and time waiting on I/O. They are not a promise of 410 hours of an arbitrary always-on server. The backend also includes an AI Gateway, but it uses purchased prepaid credits; general availability does not create free model inference.

A temporary Postgres database without an account

Neon added another useful route on 11 September: an agent can create a claimable project before its owner signs up.

Claimable Neon provides a temporary Postgres project with 100 MB storage and 1 GB transfer. The project expires after 72 hours, while its claim code expires after 15 minutes. Those are different deadlines; a project still existing does not mean its original claim link remains usable.

The Data API and Auth can be requested, but Functions, Object Storage and AI Gateway require claiming first. That makes it useful for generating and testing a database-backed prototype before onboarding. It does not replace the permanent Free plan or a backup.

Vercel sandboxes can keep a workspace between runs

Vercel Sandbox Drives entered public beta on 23 September, including Hobby. A Drive is a persistent directory that can be mounted across different sandbox runs. Files under that mount survive the sandbox stopping, so dependencies, datasets or an agent’s working directory need not be rebuilt every time.

Hobby includes 15 GB of Drive storage, plus 30 GB of reads and 30 GB of writes per month. The pricing table labels the storage allowance as lifetime capacity, separate from the monthly I/O budgets. A Drive must be mounted in its own region and supports one read-write mount at a time.

The underlying Hobby compute allowance remains five active CPU-hours, 420 GB-hours of memory and 5,000 creations per month, with ten concurrent sandboxes and a 45-minute maximum per session. Stop/resume can preserve state and start another session, but it does not replenish the monthly compute pool. New creations pause when the quota is exhausted. Hobby is for personal, non-commercial use.

Persistence itself became generally available in May. The September addition is reusable Drives, rather than a newly unlimited free VM.

There is a separate hosting change to keep in mind. Vercel tightened Hobby deployment retention on 16 September. Teams have 10 GB of deployment storage, and unprotected deployments can be deleted immediately when that limit is exceeded. The recent-deployment exceptions preserve the latest three deployments of any type and the latest three Ready production deployments; those sets can overlap. An old preview URL is a poor place to archive a demo.

Cloud Run gives patient batch jobs another option

Google added delayed job execution in Preview on 8 September. A non-urgent job can wait up to 12 hours for provisioning at a lower price. The execution can then run for up to 12 hours, so this is a possible fit for overnight processing that can tolerate an uncertain start.

Cloud Run’s pricing table lists a free monthly equivalent of 342,857 vCPU-seconds and 642,857 GiB-seconds for delayed jobs at Tier 1 pricing. The qualification matters more than another pair of large numbers: Google aggregates the free tier by billing account and applies it as a spending discount.

Do not add the delayed-job, ordinary-job and request-based rows together as independent grants. The account is billing-enabled, and builds, stored images, networking and excess usage can still charge. The delayed-job documentation explains the scheduling limits before you build a batch workflow around it.

Three credit and snapshot corrections

Modal still provides $30 of compute each month, but its hosted API is a separate bill. The pricing FAQ explicitly says Shared Endpoints tokens are charged from the first request and excluded from Starter’s compute credit, effective 1 September 2026. The credit covers Functions, Sandboxes and Notebooks metered for CPU, GPU and memory. Set a Workspace spend limit for the cash cap; a usage budget is measured before credits.

Lightning’s current GPU offer is starter credit. Its live pricing page gives five credits on registration and 25 more after adding a payment card. Unused credits expire after 12 months. September described 15 recurring monthly credits; that is not supported by the current offer. One CPU Studio can remain free with a restart every four hours, but GPU work spends the credit balance. No dated announcement was found, so this is a correction to the current reference, not a claim about when Lightning changed its plan.

Freestyle includes ten ordinary snapshots. Both its pricing page and detailed limits support that figure, correcting September’s blanket “no snapshots” description. The recurring Free pool remains 200 allocated vCPU-hours, 400 GiB memory-hours and 60,000 GiB storage-hours per month, with hard stops. However, the pricing table excludes persistent Free VMs while the detailed documentation includes them. Until those pages agree, a persistent Free machine remains an unresolved promise.

Mistral publishes a monthly API credit

Mistral’s current Free pricing now gives a concrete headline: $10 per month in API credits. September’s description was more opaque and account-specific. No dated launch was established for this offer.

The subscription documentation says usage is shared across Studio, API and Vibe Code. With pay-as-you-go disabled, exhausted usage can stop until renewal; with it enabled, additional tokens can bill. Model availability and rate limits still need checking for the actual account.

The core shortlist still holds

These are the main choices rechecked for October, with the boundary kept next to the allowance.

• A Linux VM you control: OCI Always Free: 1,500 A1 OCPU-hours and 9,000 GB-hours monthly, equivalent to 2 OCPUs/12 GB RAM, plus 200 GB boot/block storage. Home-region capacity can be unavailable; idle instances can be reclaimed.

• A small API or webhook: Workers Free: 100,000 requests/day, 10 ms CPU per HTTP request, 128 MB memory. A hard request cap and a small CPU budget; storage has separate quotas.

• TypeScript application hosting: Deno Deploy: one million requests, 10 CPU-hours and 150 GiB-hours of memory monthly. Ten apps, no Free Sandbox or volume storage; September’s reduced allowance remains.

• Intermittent Postgres: Neon Free: 100 CU-hours and 0.5 GB per project. Five-minute scale-to-zero; compute/transfer exhaustion suspends compute.

• MySQL or another managed service: Aiven Free: MySQL, Postgres, Kafka, OpenSearch and Valkey. No card or fixed trial expiry, but inactive services can stop.

• Occasional GPU work: Modal’s $30 recurring compute pool. Payment method required; extra compute and Shared Endpoints can charge.

• Hosted model inference: Workers AI: 10,000 Neurons/day. Free stops at the daily allocation; some models require billing or prepaid credits.

• A temporary ARM VM promotion: AWS T4g: 750 aggregate t4g.small hours/month. Still ends 31 December 2026; disks, public IPv4 and surplus CPU can charge.

One smaller historical correction: September said GitLab’s five-user private namespace limit took effect on 15 August 2026. GitLab’s current canonical documentation dates its application to older namespaces to 13 June 2023. The five-user rule remains relevant; the August effective date should not be carried forward.

Cloudflare also rebuilt Containers for agent sandboxes on 30 September, including runtime image selection and filesystem snapshots. Its pricing still starts with the $5/month Workers Paid plan and billable resource usage. That release belongs in a sandbox comparison, but it does not add a free container allocation to Workers Free.

A credit competition with a deadline tomorrow

For an eligible startup, Cloudflare’s Cold Start competition is a timely application lead. Five finalists pitch in San Francisco on 19 October, with one winner receiving $500,000 in Cloudflare credits. It is open to early-stage companies based in the US or Canada that have raised less than $10 million, and applications close 2 October 2026.

That geographical restriction excludes a UK-only company, and the prize is competitive rather than a signup allowance. Keep the deadline attached when sharing the offer.