Free compute is easiest to misunderstand when the headline number is technically true.
A platform says it includes a million requests, five CPU-hours, or $200 of compute. That does not tell you whether usage stops at the limit, whether a card can be charged, whether the credit happens once, or whether the capacity will exist when you need it.
I rechecked the free-compute list for September against current provider documentation. This month produced fewer dramatic closures than August, but it exposed several numbers that needed correcting and a useful new pattern: infrastructure an agent can create before you even make an account.
Correct the old numbers first
Deno Deploy Free is smaller than it was in the August snapshot. The current plan lists 1 million requests, 20 GiB egress, 10 CPU-hours, 150 GiB-hours of memory, ten apps, no Sandbox or volume storage, and 1 GiB KV per month.
That replaces the previous 15 CPU-hours, 350 GiB-hours, 20 apps, and 1 GiB volume allowance. Deno does not expose a date for the change. Its changelog also warns that unverified organisations can receive restricted limits until a payment method is linked.
Four other corrections matter:
- Freestyle VMs provides 200 vCPU-hours, 400 GiB memory-hours, and 60,000 GiB storage-hours per month. Those are monthly figures, not daily ones.
- Browserless now documents two-minute sessions, alongside 1,000 monthly units and two concurrent browsers.
- Firebase Authentication does not simply provide 50,000 MAU on Spark. Identity Platform on Spark is normally limited to 3,000 DAU, while the 50,000-MAU no-cost tier belongs to Identity Platform on the billing-enabled Blaze plan.
- Axiom Personal currently includes three datasets rather than two, with 25 GB stored, 500 GB monthly ingest, and 10 GB-hours of query compute.
None of those is a glamorous new product launch. They are still more useful than carrying a wrong number into another month's list.
Know how the free usage ends
Cloudflare D1 now has an unusually clear answer. Since 1 September, a Free account that reaches 5 million rows read or 100,000 rows written in a day gets query errors until midnight UTC. Stored data remains intact.
That is a hard cap. It can break an application, but it cannot quietly become a query-overage bill.
Vercel Sandbox works the same way at the account level. Hobby includes five CPU-hours, 420 GB-hours of memory, 5,000 creations, 20 GB transfer, 15 GB of lifetime snapshots, 45-minute sessions, and ten concurrent sandboxes. Creation pauses after the allowance is exhausted.
Compare that with Google Cloud Run. Its 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds are recurring free usage inside a billing account. Build, registry, storage, network, and excess runtime can still cost money.
Both are legitimately useful. They just fail differently.
Infrastructure before signup
Railway now allows an unauthenticated visitor to create a private site through Dev New or start a database. The build session lasts 60 minutes, the result remains claimable for 24 hours, anonymous workloads can be throttled, and agent building is capped at $3 of LLM credit.
Cloudflare temporary accounts provide a similar one-hour loop from the command line. An agent can deploy a Worker with supported resources, verify it, and return a claim link. If the claim is not completed within 60 minutes, the account and resources disappear.
Axiom's agent-created organisations bring temporary observability into the same pattern. An agent can create an unauthenticated US East organisation with 10 GB ingest and one GB-hour of query compute. It must be claimed within 24 hours.
These offers are not permanent free hosting. They are disposable proof environments with a handoff step. Claim links should be treated like credentials, and every unclaimed environment is designed to vanish.
A practical shortlist
For a small real VM, OCI Always Free remains the standout if Ampere capacity exists in your home region. The supported allowance is 1,500 OCPU-hours and 9,000 GB-hours per month, equivalent to 2 OCPUs and 12 GB RAM, plus two AMD micro instances and 200 GB combined boot and block storage.
For application code, Cloudflare Workers gives 100,000 requests per day with 10 ms CPU per HTTP request. Deno remains useful if its reduced allowance fits. Cloud Run is the better fit when you need a container and are willing to manage a billing-linked account.
For data, Neon remains a strong intermittent Postgres option with 100 CU-hours per project and scale-to-zero. Turso covers distributed SQLite with 5 GB storage, 500 million rows read, and 10 million rows written per month. D1 fits naturally beside Workers now that its stopping rule is explicit.
For disposable code execution, start with Vercel Sandbox, Upstash Box, or Freestyle. For browser execution, Cloudflare Browser Run provides ten minutes per day, Browserbase provides one hour per month, and Browserless provides short sessions from its unit pool.
For burst compute, Modal still provides $30 of recurring credit across CPU, GPU, notebooks, schedules, sandboxes, and web functions. Paperspace still lists capacity-dependent C4 and M4000 notebooks. Hugging Face ZeroGPU is much narrower: eligible personal accounts receive five shared GPU minutes per day for public Gradio Spaces.
Hosted model APIs remain harder to compare. Workers AI gives 10,000 Neurons per day, but some models require billing or prepaid credits. Gemini's free limits vary by model, project, account, and region. OpenRouter provides 50 free-model requests per day, rising to 1,000 after at least $10 of credit has been purchased.
The useful check is not just “how much is free?” Ask what happens at the edge: stop, sleep, delete, throttle, expire, or bill. September's changes make several of those answers clearer.
Want to talk about something I’ve written or built? Get in touch.
This article was adapted with AI assistance from an original article on billiem.uk. The original article was reviewed before publication.
Top comments (0)