A monthly free-compute list is only useful if it is willing to delete things.
Between my July snapshot and 3 August, GitHub Models disappeared, SageMaker Studio Lab closed to new users, and the current Hugging Face documentation stopped supporting the broad free CPU Spaces claim I had made. Paperspace, meanwhile, turned up with documented free M4000 notebook access. Neon put Functions and Object Storage into a free beta. Managed agent runtime and browser automation became harder to treat as side notes.
I also had to correct my own July guide. GitHub had already announced that Models would retire on 30 July, but I still described it as a recurring public preview on 10 July. That was wrong.
This is the useful August shortlist: what to remove, what changed, and which kind of free allowance fits a particular side project. It is documentation-backed rather than a claim that I opened and tested every account. Capacity, regions, billing checks, data-use terms, and unpublished guardrails still matter.
Remove these from the old shortlist
GitHub Models retired completely, including its playground, catalogue, inference API, and BYOK endpoints. It is no longer a free model API.
SageMaker Studio Lab closed to new customers on 30 July. Existing users retain free sessions with 15 GB storage, 16 GB RAM, up to eight CPU hours or four GPU hours per day, subject to capacity. A new reader cannot claim it.
Hugging Face Spaces needs a narrower description. Normal Gradio and Docker compute Spaces are gated behind paid plans in the current documentation. A free personal account can still host up to two Gradio ZeroGPU Spaces, but that is queued public-demo GPU capacity rather than a free Linux box.
One more correction is smaller but worth keeping: Axiom Personal currently includes two datasets, not the three I listed in July.
The August additions worth knowing about
A genuinely free GPU notebook appeared
Paperspace Notebooks now documents C4 CPU and M4000 GPU machines on its Free Gradient plan. It includes 5 GB storage and permits one running notebook at a time. A free-machine session can last up to six hours.
That does not mean guaranteed GPU hours. Capacity may queue, and there is no published monthly-hour entitlement. It is useful shared notebook access, not a persistent GPU server.
Neon temporarily became more than Postgres
Neon Functions and Object Storage are available on the Free plan during beta. The current offer is restricted to us-east-2, retains logs for three days, and uses unpublished rate and usage guardrails. AI Gateway is not included on Free.
That makes Neon an interesting temporary full-stack backend, but a beta is not a promise. The underlying free Postgres plan remains easier to reason about: up to 100 projects with 100 CU-hours and 0.5 GB storage per project, plus scale-to-zero after five minutes.
Agent compute now has several shapes
Google Agent Engine is listed with 180,000 vCPU-seconds and 360,000 GiB-seconds per month. Billing must be enabled. The table does not establish that Code Execution sandboxes share this allowance, so I would use the claim narrowly: managed agent runtime.
Upstash Box provides ten concurrent boxes and five active CPU-hours each month. A box has 2 vCPU, 4 GB RAM, and 5 GB disk, freezes after an idle hour, and currently runs only in us-east-1.
Cloudflare temporary accounts solve a different problem. An unauthenticated agent can deploy a Worker and supported resources with wrangler deploy --temporary, verify the result, and return a claim URL. The account disappears after 60 minutes if it is not claimed. This is throwaway deployment compute, not a permanent anonymous account.
Browser automation is now its own free-compute category
Cloudflare Browser Run gives ten browser minutes per day and three concurrent browsers. Cloudflare's general-purpose Sandbox SDK is separate and has no Workers Free allocation.
Browserbase includes one browser hour, three concurrent browsers, three agent runs, and 1,000 each of Search and Fetch uses per month. Sessions stop after 15 minutes.
Browserless uses a 1,000-unit monthly pool with two concurrent browsers and one-minute sessions. One unit covers up to 30 seconds, while proxy and CAPTCHA work can consume more. It does not translate cleanly into a neat number of hours.
Pick by the way the allowance can fail
The word “free” hides several different products. The failure mode matters more than the headline number.
For a persistent Linux VM
OCI Always Free is still the first claim to try. The current supported allocation is 1,500 Ampere A1 OCPU-hours and 9,000 GB-hours per month, equivalent to 2 OCPUs and 12 GB RAM. The tenancy also gets 200 GB combined boot and block storage plus 10 TB monthly outbound transfer.
The catches are practical: home-region capacity can be unavailable, and idle instances can be reclaimed. AWS EC2 T4g is useful during 2026, but its 750 aggregate t4g.small hours per month are a promotion ending on 31 December rather than durable infrastructure.
For functions and containers
Cloudflare Workers has a recurring hard cap of 100,000 requests per day, 10 ms CPU per HTTP request, and 128 MB memory. Static assets do not consume the dynamic request allowance.
Deno Deploy publishes a broader monthly cap: 1 million requests, 20 GB egress, 15 CPU-hours, 350 GB-hours memory, 20 apps, 1 GiB volume storage, and 1 GiB KV storage. The old Deploy Classic migration deadline has passed, and projects did not transfer automatically.
Google Cloud Run gives request-based services 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds monthly. It is attached to billing, so registry, build, network, storage, or overage usage can still charge.
For data and product plumbing
Turso is the cleaner hard-capped SQLite-shaped option: 100 databases, 5 GB storage, 500 million rows read, and 10 million rows written per month. Requests stop at quota rather than becoming an invoice.
Cloudflare D1 fits naturally when the application already lives on Workers. The free allocation is ten databases, 500 MB each, 5 GB total, 5 million rows read and 100,000 rows written per day.
Upstash covers Redis, QStash, Workflow, Vector, and preview Search under separate free quotas. Cloudflare R2 provides 10 GB-month Standard storage, 1 million Class A operations, 10 million Class B operations, and free egress, but enabling it requires account checkout.
For GPU and hosted inference
Modal Starter gives $30 of recurring compute credit for serverless CPU, GPU, notebooks, schedules, sandboxes, and web functions. Lightning AI Free gives 15 credits per month plus one active CPU Studio that restarts every four hours. The GPU time those credits buy varies with the selected hardware and current rates.
Google Colab remains shared and dynamic, with no stable hardware or quota promise. Sessions can run for up to 12 hours depending on usage and availability. Paperspace is now the more concrete free M4000 notebook claim, while ZeroGPU is for public Gradio demos.
Hosted APIs are another category again. Cloudflare Workers AI stops a Free account after 10,000 Neurons per day. Gemini has model-dependent free pricing and may use free-tier content to improve Google products. OpenRouter Free gives 50 free-model requests per day unless the account has purchased credit.
FreeInference is useful only with its privacy boundary visible: its terms allow prompts and responses to be logged and anonymised derived data to be released. Never send it secrets, private code, or personal data.
Three stacks I would still assemble from the list
For a small real server: OCI for the VM, Cloudflare for DNS, CDN, and TLS, R2 or Tigris for objects, Neon or Aiven for Postgres, Upstash for queues and Redis, Resend for mail, and Grafana Cloud plus Sentry for visibility.
For a no-server product: Cloudflare Workers and static assets, D1 or Neon, R2, Queues or Workflows, Turnstile, Clerk or AuthKit, and Workers AI.
For an AI demo: Paperspace or Colab for notebook work, Modal for repeatable bursts, ZeroGPU for a public Gradio surface, and Workers AI, Gemini, Groq, or OpenRouter for hosted inference.
The full August guide on billiem.uk keeps the broader directory, including email, auth, observability, CI, storage, startup programmes, and academic credits.
In 24 days, one model API disappeared, one notebook closed to new users, and a supposed free hosting offer became much narrower. The next version may remove as much as it adds. That is fine. A dated guide is more useful when it is willing to get shorter.
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)