Free compute worth claiming in September 2026

A useful zero-cost stack is rarely one provider's free plan. It is usually a collection of services with completely different failure modes.

One component stops at a daily cap. Another scales to zero. A third can charge the card attached to the account. A notebook may be free only when capacity is available, while a startup programme is not available at all until somebody approves the application.

The September free-compute check found fewer closures than August, but it made those boundaries clearer. It also found a new disposable stack that can be deployed, observed, and handed over before the owner has created permanent accounts.

A disposable first deployment

Railway's new anonymous path can create a private site through Dev New or start a database without an account. The build session lasts 60 minutes. The result remains available to claim for 24 hours, capacity can be throttled, and agent building is limited to $3 of LLM credit.

Cloudflare temporary accounts cover a similar command-line loop. An agent can deploy a Worker with supported KV, D1, Durable Objects, Hyperdrive, or Queues resources and return a claim URL. The claim must be completed within 60 minutes or the temporary account and its resources are deleted.

Axiom can now create an unauthenticated organisation for an agent. The temporary US East organisation has 10 GB ingest and one GB-hour of query compute. Notifiers are disabled and the organisation disappears permanently if it is not claimed within 24 hours.

Together, these are enough for an agent to build a small application, put it online privately, collect logs, and hand the whole experiment back. They are not a production stack. The deletion clock is part of the product, and the claim links are credentials in practice.

A small application that should stop rather than bill

Cloudflare Workers remains a sensible compute layer when 100,000 daily requests and 10 ms CPU per HTTP request are enough. Static assets sit outside the dynamic request allowance.

Cloudflare D1 fits beside it with ten databases, 5 GB total storage, 5 million rows read per day, and 100,000 rows written. The important September change is not the quota. Since 1 September, queries fail after the daily row limit is reached and resume at midnight UTC. Stored data stays intact.

Add R2 for 10 GB-month of Standard storage, one million Class A operations, ten million Class B operations, and free egress. R2 does require checkout and can bill overages, so it does not share D1's failure behaviour even though both sit in the same stack.

For queues or durable jobs, Workers Free includes 10,000 Queue operations per day and Workflows includes 3,000 steps per day plus 1 GB-month state storage. A normal delivered Queue message commonly consumes send, receive, and acknowledgement operations, so the message count is lower than the operation count.

That is a useful no-server product shape, but only if the R2 billing boundary is treated separately from the hard-capped pieces.

A small application that needs Postgres

Neon remains the easiest free Postgres recommendation for intermittent work. Each Free project receives 100 CU-hours per month, 0.5 GB storage, compute up to 2 CU and 8 GB RAM, and five-minute scale-to-zero.

Its Functions and Object Storage products are free during beta in us-east-2, with unpublished guardrails. That can be useful for an experiment, but it is not the same promise as the quantified database allowance.

Aiven's Free tier now spans PostgreSQL, MySQL, Kafka, OpenSearch, and Valkey. The services are small, may power off while idle, and have no SLA. Turso is the alternative when distributed SQLite is the better architecture: 100 databases, 5 GB storage, 500 million rows read, 10 million rows written, and 3 GB embedded sync per month.

For vectors, Qdrant Cloud provides 1 GB RAM, 0.5 vCPU, and 4 GB disk. It suspends after an idle week and can be deleted after four. Pinecone Starter provides five serverless indexes and 2 GB storage in AWS us-east-1, with separate read, write, and embedding allowances.

Disposable compute with numbers attached

Vercel Sandbox now publishes a useful Hobby allowance: five CPU-hours, 420 GB-hours memory, 5,000 creations, 20 GB transfer, 15 GB of lifetime snapshots, 45-minute sessions, and ten concurrent sandboxes. New creations pause after the quota is exhausted.

Upstash Box gives ten concurrent boxes and five active CPU-hours per month. Each box has 2 vCPU, 4 GB RAM, and 5 GB disk in us-east-1. Freestyle VMs publishes 200 vCPU-hours, 400 GiB memory-hours, and 60,000 GiB storage-hours per month, with ten concurrent or saved VMs but no persistent VMs or snapshots.

That corrects an error in the August guide, where Freestyle's monthly allowance was described as daily.

Browser compute has smaller clocks. Cloudflare Browser Run includes ten minutes per day and three concurrent browsers. Browserbase includes one browser hour, three agent runs, and 1,000 each of Search and Fetch per month. Browserless includes 1,000 units, two concurrent browsers, and two-minute sessions. August reported one-minute sessions; the current pricing page says two and does not date the change.

Compute for model work

Modal remains the clearest recurring burst-compute offer: $30 per month that can pay for CPU, GPU, notebooks, schedules, sandboxes, and web functions. The credit is stable even though the GPU time it buys depends on the selected hardware.

Paperspace continues to list free C4 CPU and M4000 GPU notebook machines, 5 GB storage, and six-hour maximum Free sessions. Capacity can queue. Its current pricing page says the machines run in private workspaces, while an older official page says free Gradient machines are always public, so sensitive data needs a check in the actual workspace rather than trust in either sentence.

Hugging Face ZeroGPU is now quantifiable. A personal account needs a verified email, more than 30 days of age, and good standing. It can host two Gradio ZeroGPU Spaces and receives five GPU minutes per day. That is enough for a small public demo, not a free training machine.

For hosted inference, Workers AI provides 10,000 Neurons per day but reserves some models for billing-enabled or prepaid accounts. Gemini's free limits vary by model, account, project, and region. OpenRouter provides 50 free-model requests per day without purchased credit. Cerebras is now a $5 one-time trial, so it stays outside the recurring list.

The corrections matter more than the catalogue

Deno Deploy Free now lists 10 CPU-hours, 150 GiB-hours memory, ten apps, no Sandbox or volume storage, and 1 GiB KV. August's higher CPU, memory, app, and volume figures are stale.

Firebase Authentication also needed a sharper boundary. Identity Platform on Spark is normally limited to 3,000 DAU. The 50,000-MAU no-cost tier applies after upgrading Identity Platform to the billing-enabled Blaze plan.

Axiom Personal has three datasets rather than the two reported last month. Microsoft's startup documentation currently disagrees with itself about whether the milestone path starts with $200 or $1,000, so neither figure belongs in a confident recommendation.

This is what a monthly free-compute article is for. The useful result is not the longest list. It is knowing which component will stop, sleep, disappear, throttle, require approval, or reach a bill when the free part ends.

---

Want to talk about something I’ve written or built? Get in touch.

This Hashnode edition was adapted from the original article on billiem.uk, which remains the canonical and most complete version.