--- title: "Free compute worth claiming in September 2026" date: 2026-09-04 description: "A September 2026 guide to changed free tiers, corrected limits, disposable sandboxes, VPSes, databases, AI APIs, and GPU notebooks." tags: - "self-hosting" - "free-tier" - "side-projects" - "cloud-infrastructure" - "ai-tools" --- This is the 4 September 2026 snapshot of free infrastructure worth claiming for side projects. It follows the [August guide](/posts/free-compute-august-2026/), but a monthly free-compute article does not need to pretend every month produces a completely new cloud market. September is mostly a month of sharper limits. Deno Deploy has cut several Free allowances. Cloudflare D1 now actually stops queries when the daily quota is exhausted. A few agent sandboxes have finally published useful numbers. I also need to correct four things from August: Freestyle's allowance is monthly rather than daily, Browserless sessions are now two minutes, Firebase's 50,000-user claim needed a billing-plan qualification, and Axiom has three datasets rather than two. There is one genuinely fun addition. Railway now lets someone, or an agent, deploy a private site or database for an hour without creating an account. It is not permanent free hosting, but it joins Cloudflare and Axiom in a small category of infrastructure that can exist before the signup form. The figures below were checked against official provider pages on 3 September. This check did not create and test every account, and documentation does not prove that capacity will exist in your region. Where a provider's own pages conflict, the conflict stays visible or the number stays out. ## What changed for September - [Deno Deploy Free](https://deno.com/deploy/pricing) now 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. August listed 15 CPU-hours, 350 GiB-hours, 20 apps, and 1 GiB of volume storage. The current pricing table is the one to use, although Deno does not publish a date for the reduction. Its [changelog](https://docs.deno.com/deploy/changelog/) also says unverified organisations can receive restricted limits until a payment method is linked, and apps pause when they exceed quota. - [Cloudflare D1 began enforcing its Free daily query limits on 1 September](https://developers.cloudflare.com/changelog/post/2026-09-01-d1-free-tier-limit-enforcement/). The allowance remains 5 million rows read and 100,000 rows written per day, but queries now fail until midnight UTC after either limit is reached. Stored data is unaffected. That makes D1 a real hard cap rather than a vague free allowance with ambiguous overage behaviour. - [Railway's anonymous deployment preview](https://blog.railway.com/p/deploy-without-account) gives an unauthenticated visitor 60 minutes to build a private Railway Site through Dev New or create a database. The result remains claimable for 24 hours. Railway can throttle anonymous deployments when demand is high, and agent building is limited to $3 of LLM credit. This is a try-before-signup path, not a replacement for Railway's normal trial or Free plan. - [Vercel Sandbox](https://vercel.com/docs/sandbox/pricing) now gives Hobby accounts a properly documented recurring allowance: five CPU-hours, 420 GB-hours of memory, 5,000 sandbox creations, 20 GB transfer, 15 GB of lifetime snapshots, 45-minute sessions, and ten concurrent sandboxes. New sandboxes pause when the quota is exhausted rather than creating overage charges. - [Freestyle VMs](https://www.freestyle.sh/pricing) provides 200 vCPU-hours, 400 GiB memory-hours, and 60,000 GiB storage-hours per month. I called those daily allowances in August. That was wrong. Free also allows ten concurrent or saved VMs, but not persistent VMs or snapshots. - [Browserless](https://www.browserless.io/pricing) still provides 1,000 units per month and two concurrent browsers without a card. Its current maximum session length is two minutes, not the one minute I reported in August. The provider does not date the change. - [Hugging Face ZeroGPU](https://huggingface.co/docs/hub/main/en/spaces-zerogpu) is narrower and more measurable than the August description. A free personal account must have a verified email, be more than 30 days old, and remain in good standing. It can host two Gradio ZeroGPU Spaces and receives five GPU minutes per day. This is shared capacity for public demos, not a free GPU notebook or Linux machine. - [Firebase Authentication's limits](https://firebase.google.com/docs/auth/limits) need a correction. Identity Platform on Spark is normally limited to 3,000 daily active users, or two daily active users for SAML and OIDC. The 50,000 monthly active user no-cost tier applies after upgrading to Identity Platform on the billing-enabled Blaze plan. August collapsed those into one misleading sentence. - [Axiom Personal](https://axiom.co/docs/reference/limits) currently allows three datasets, not two, alongside 25 GB stored, 500 GB loaded per month, 10 GB-hours of query compute, and 30-day retention. Axiom has also added [agent-created organisations](https://axiom.co/docs/console/intelligence/agent-created-orgs): an agent can provision a temporary US East organisation without authentication, use 10 GB ingest and one GB-hour of query compute, then hand back a claim link. It is deleted permanently if nobody claims it within 24 hours. - Google's free-tier table now calls Agent Engine [Agent Runtime on Gemini Enterprise Agent Platform](https://docs.cloud.google.com/free/docs/free-cloud-features). The recurring billing-linked allowance remains 180,000 vCPU-seconds and 360,000 GiB-seconds per month. The renamed product should not be confused with a free Gemini API allowance or a general code-execution sandbox. ## The short version - For a real Linux VM, [OCI Always Free](https://docs.oracle.com/en-us/iaas/Content/FreeTier/freetier_topic-Always_Free_Resources.htm) remains the largest durable offer if capacity exists in your home region. Google's `e2-micro` is smaller and billing-linked. AWS's `t4g.small` promotion is useful only until 31 December 2026. - For small APIs, use [Cloudflare Workers](https://developers.cloudflare.com/workers/platform/limits/) when ten milliseconds of CPU is enough, [Deno Deploy](https://deno.com/deploy/pricing) when its reduced TypeScript-first allowance fits, or [Google Cloud Run](https://cloud.google.com/run/pricing) when you need a container and accept billing-linked usage. - For intermittent Postgres, [Neon](https://neon.com/pricing) still gives each Free project 100 CU-hours and scales to zero. Use [Turso](https://turso.tech/pricing) when distributed SQLite is the better fit, or D1 when the application already lives on Workers and a daily hard stop is useful. - For disposable agent compute, start with Vercel Sandbox, [Upstash Box](https://upstash.com/pricing/box), or Freestyle VMs. Railway and [Cloudflare temporary accounts](https://developers.cloudflare.com/workers/platform/claim-deployments/) now cover the even earlier stage where an agent needs to deploy and verify something before the owner has signed up. - For burst GPU or CPU jobs, [Modal](https://modal.com/pricing) still provides $30 of recurring compute credit. Paperspace offers capacity-dependent M4000 notebooks. ZeroGPU is for small public Gradio demos, not development sessions. - Anything attached to AWS, Azure, Google Cloud, Cloudflare R2, or another billing profile still needs a budget alert and a deliberate shutdown rule. A free allowance inside a billable account is not the same thing as a hard-capped Free plan. ## VMs, containers, and application platforms - [OCI Always Free](https://docs.oracle.com/en-us/iaas/Content/FreeTier/freetier_topic-Always_Free_Resources.htm): **Always free, capacity-dependent.** Up to two AMD micro VMs, plus 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 and 10 TB monthly outbound transfer. Compute must live in the home region, capacity can be unavailable, and idle instances can be reclaimed. - [Google Compute Engine](https://docs.cloud.google.com/free/docs/free-cloud-features): **Billing-linked recurring allowance.** One non-preemptible `e2-micro` instance-month, 30 GB-month standard persistent disk, and 1 GB outbound per month in Oregon, Iowa, or South Carolina. A billing account is required and excess or adjacent usage can charge. - [AWS EC2 T4g](https://aws.amazon.com/ec2/instance-types/t4/): **Promotion through 31 December 2026.** The offer covers 750 aggregate `t4g.small` hours per month. EBS, public IPv4, non-covered images, and surplus CPU can still cost money. - [Cloudflare Workers](https://developers.cloudflare.com/workers/platform/limits/) and [Workflows](https://developers.cloudflare.com/workflows/reference/pricing/): **Recurring hard caps.** Workers includes 100,000 requests per day, 10 ms CPU per HTTP request, and 128 MB memory. Workflows separately includes 3,000 steps per day and 1 GB state storage. Free does not bill step or storage overages. - [Google Cloud Run](https://cloud.google.com/run/pricing): **Billing-linked recurring allowance.** Request-based services get 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds per month. Builds, images, storage, networking, and usage outside the allowance can still charge. - [Azure Functions](https://azure.microsoft.com/en-us/pricing/details/functions/) and [Azure Container Apps](https://azure.microsoft.com/en-us/pricing/details/container-apps/): **Billing-linked recurring allowances.** Classic Functions includes 1 million executions and 400,000 GB-seconds. Container Apps includes 2 million requests, 180,000 vCPU-seconds, and 360,000 GiB-seconds. Storage, network traffic, and minimum replicas sit outside the grant. - [Vercel Hobby](https://vercel.com/pricing), [Netlify Free](https://docs.netlify.com/manage/accounts-and-billing/billing/billing-for-credit-based-plans/how-credits-work/), [Render Free](https://render.com/docs/free), and [Railway Free](https://docs.railway.com/pricing/free-trial): **Small application allowances with very different boundaries.** Vercel Hobby is personal and non-commercial. Netlify shares 300 monthly credits across deploys, compute, bandwidth, agent runners, and AI, then pauses every project. Render gives 750 instance-hours per workspace but sleeps idle services and treats them as non-production. Railway begins with a $5, 30-day trial, then falls back to $1 of monthly credit and can reject Free deployments during regional peaks. - [AWS Free plan](https://docs.aws.amazon.com/awsaccountbilling/latest/aboutv2/free-tier-plans.html), [Azure free account](https://azure.microsoft.com/en-us/free/), [Google Cloud trial](https://docs.cloud.google.com/free/docs/free-cloud-features), [DigitalOcean](https://www.digitalocean.com/pricing), [Fly.io](https://fly.io/docs/about/free-trial/), [Akamai Cloud](https://www.akamai.com/lp/free-credit-100), [OVHcloud](https://www.ovhcloud.com/en-gb/public-cloud/free-trial/), and [Vultr](https://docs.vultr.com/support/platform/billing/how-can-i-add-a-promotional-code): **Trials and credits.** These are useful for evaluation or a short burst, but they are not permanent free VPSes. DigitalOcean's current offer is $5 for 90 days, not the older $200 headline that still appears in stale lists. ## Sandboxes and browser compute - [Cloudflare temporary accounts](https://developers.cloudflare.com/workers/platform/claim-deployments/): **One-hour deployment preview.** `wrangler deploy --temporary` can create a Worker and supported resources, including KV, one small D1 database, Durable Objects, Hyperdrive, and Queues. Opening the claim link is not enough: the claim flow must be completed within 60 minutes or the account and resources are deleted. - Railway's anonymous preview: **One-hour private build session.** It is easiest through Dev New, remains claimable for 24 hours, and is capacity-dependent. Unlike Cloudflare's path, the unclaimed workload is not immediately public. - Vercel Sandbox, Upstash Box, and Freestyle VMs: **Recurring disposable compute.** Vercel supplies the most clearly documented general allowance. Upstash gives ten concurrent 2 vCPU, 4 GB boxes and five active CPU-hours per month in `us-east-1`, plus $1 monthly built-in agent-token allowance. Freestyle publishes much larger monthly CPU, memory, and storage figures, but excludes persistent VMs and snapshots. - [Cloudflare Browser Run](https://developers.cloudflare.com/browser-run/pricing/), [Browserbase](https://www.browserbase.com/pricing), and Browserless: **Recurring browser compute.** Browser Run provides ten minutes per day and three concurrent browsers. Browserbase provides one browser hour, three agent runs, and 1,000 each of Search and Fetch per month with 15-minute sessions. Browserless uses a 1,000-unit pool, two concurrent browsers, and two-minute sessions; proxy and CAPTCHA work can consume extra units. - [E2B](https://e2b.dev/pricing) and [Daytona](https://www.daytona.io/pricing): **Starter credits with incomplete recurrence evidence.** E2B's $100 Hobby credit is explicitly one-time. Daytona still advertises $200 without publishing a clean recurrence or expiry rule, so it stays outside the recurring shortlist. ## Databases, storage, and product plumbing - Neon: **Recurring, no card for the base database plan.** Up to 100 projects, each with 100 CU-hours, 0.5 GB storage, compute up to 2 CU and 8 GB RAM, and five-minute scale-to-zero. Functions and Object Storage remain free during beta in `us-east-2`, with unpublished guardrails. - Turso and D1: **Hard-capped databases.** Turso includes 100 databases, 5 GB storage, 500 million rows read, 10 million rows written, and 3 GB embedded sync per month. D1 includes ten databases, 5 GB total, and the newly enforced daily row limits. - [Aiven](https://aiven.io/docs/platform/concepts/service-pricing): **Recurring shared services.** Its Free tier now covers PostgreSQL, MySQL, Kafka, OpenSearch, and Valkey. The small services can power off while idle and do not carry an SLA. - [Qdrant Cloud](https://qdrant.tech/documentation/cloud/create-cluster/) and [Pinecone Starter](https://docs.pinecone.io/reference/api/database-limits): **Recurring vector stores.** Qdrant's free cluster has 1 GB RAM, 0.5 vCPU, and 4 GB disk; it suspends after an idle week and can be deleted after four. Pinecone gives five serverless indexes and 2 GB storage in AWS `us-east-1`, with monthly read, write, and embedding allowances. - [Cloudflare R2](https://developers.cloudflare.com/r2/pricing/), [Tigris](https://www.tigrisdata.com/pricing/), and [Backblaze B2](https://www.backblaze.com/cloud-storage/pricing): **Recurring object storage.** R2 gives 10 GB-month Standard storage, one million Class A operations, ten million Class B operations, and free egress, but enabling it requires checkout and overages can bill. Tigris provides 5 GB with small operation allowances and free egress. B2 provides the first 10 GB plus free monthly egress up to three times average storage. - [R2 Data Catalog](https://developers.cloudflare.com/r2-data-catalog/platform/pricing/): **Specialist billing-linked allowance.** It includes one million catalogue operations, 10 GB of compaction data, and one million compaction objects per month. Billing began for non-enterprise accounts on 3 August, so overages belong in a real budget even though the opening allocation is free. - [Resend](https://resend.com/docs/knowledge-base/what-is-resend-pricing), [Clerk](https://clerk.com/pricing), [WorkOS AuthKit](https://workos.com/pricing), [Svix](https://www.svix.com/pricing/), and [Hookdeck](https://hookdeck.com/pricing): **Recurring application plumbing.** The useful headline numbers remain 3,000 monthly emails for Resend, 50,000 retained users per Clerk app, one million AuthKit users with billing details for production, 50,000 attempted Svix messages, and 10,000 Hookdeck events. ## Observability, CI, and workspaces - [Grafana Cloud](https://grafana.com/pricing/?tab=free), [Better Stack](https://betterstack.com/pricing), [Sentry](https://sentry.io/pricing/), Axiom, [Honeycomb](https://www.honeycomb.io/pricing), and [New Relic](https://newrelic.com/pricing/): **Recurring observability.** Their free plans remain useful, but the products meter different things. Pick from the signal you actually need rather than comparing one headline quota. Axiom is unusually generous on ingestion; Sentry is easier to reason about for application errors and traces; Better Stack combines small monitoring and telemetry allowances. - [PostHog](https://posthog.com/) and [Flagsmith](https://www.flagsmith.com/pricing): **Recurring product telemetry.** PostHog includes one million analytics events, 5,000 replays, and one million feature-flag requests. Flagsmith includes 50,000 flag requests for one member and one project, then blocks Free requests after its warning period rather than billing an overage. - [GitHub Actions and Codespaces](https://docs.github.com/en/billing/reference/product-usage-included): **Recurring developer compute.** Standard Actions runners are free for public repositories; GitHub Free includes 2,000 private-repository minutes. Personal accounts get 120 Codespaces core-hours and 15 GB-month storage. - [GitLab CI](https://docs.gitlab.com/ci/pipelines/compute_minutes/), [CircleCI](https://circleci.com/pricing/), and [Google Cloud Build](https://cloud.google.com/build/pricing): **Recurring CI allowances.** GitLab includes 400 compute minutes and its five-user rule for private Free top-level namespaces has been effective since 15 August. CircleCI includes 30,000 credits, roughly 6,000 small Docker minutes. Cloud Build includes 2,500 `e2-standard-2` build minutes after billing is linked. ## GPU notebooks and model APIs - Modal: **Recurring $30 compute credit.** It remains the cleanest general burst-compute offer in this list because the same credit can pay for CPU, GPU, notebooks, schedules, sandboxes, and web functions. The amount of GPU time changes with the selected hardware. - [Paperspace](https://docs.digitalocean.com/products/paperspace/pricing/): **Shared free notebooks.** C4 CPU and M4000 GPU machines, 5 GB storage, and six-hour maximum Free sessions remain documented. Capacity can queue. Current pricing says Free machines are for Notebooks in private workspaces, while an older official page says Free Gradient machines are always public, so do not assume privacy without checking the actual workspace. - [Lightning AI](https://lightning.ai/pricing/): **Recurring credits and shared CPU.** Free remains 15 credits per month plus one 4-CPU Studio that restarts every four hours. Phone verification, country availability, and waitlisting can apply. Lightning's own pages disagree on the GPU hours those credits represent, so quote the credits rather than a promise of runtime. - [Google Colab](https://research.google.com/colaboratory/faq.html) and [Databricks Free Edition](https://docs.databricks.com/aws/en/getting-started/free-edition-limitations): **Dynamic notebook compute.** Colab publishes no stable free GPU quota. Databricks provides non-commercial serverless notebooks and limited capacity-dependent GPU access, with restricted outbound networking and terms that permit training on Free Edition data. - ZeroGPU: **Five shared GPU minutes per day for eligible public Gradio Spaces.** That is enough for a tiny public demonstration. It is not a hidden replacement for the broad free CPU Spaces claim removed in August. - [Cloudflare Workers AI](https://developers.cloudflare.com/workers-ai/platform/pricing/): **Recurring hard cap for eligible models.** Free includes 10,000 Neurons per day and stops when the daily allocation is empty. Some frontier models require a billing method or prepaid AI Gateway credits, so the quota does not imply that every listed model is free. - [Gemini Developer API](https://ai.google.dev/gemini-api/docs/pricing), [GroqCloud](https://console.groq.com/docs/rate-limits), and [OpenRouter](https://openrouter.ai/docs/api_reference/limits): **Model-dependent or rate-limited inference.** Gemini's free pricing varies by model, account, project, and region, and free-tier content may be used to improve Google products. Groq's limits are model- and organisation-specific, so there is no single Free table worth quoting. OpenRouter allows 50 free-model requests per day, rising to 1,000 after at least $10 of credit has been purchased. - [Mistral](https://docs.mistral.ai/admin/billing-usage/subscriptions), [FreeInference](https://freeinference.org/terms), and [Hugging Face Inference Providers](https://huggingface.co/docs/inference-providers/en/pricing): **Small or opaque evaluation access.** Mistral's limits are account-specific. FreeInference has unpublished quota and unusually permissive prompt and response logging terms. Hugging Face gives Free users $0.10 of routed inference credit per month, explicitly subject to change. - [Cerebras](https://www.cerebras.ai/pricing): **One-time trial.** Its current offer is $5 trial credit, followed by a minimum $10 payment for developer usage. It does not belong in the recurring API shortlist. ## Credits that require a qualifying project The headline figures in this category can dwarf normal free tiers, but none is infrastructure that appears merely because a signup form was completed. - [AWS Activate](https://aws.amazon.com/startups/credits/) and [Google for Startups AI](https://cloud.google.com/startup/ai): application and affiliation-based cloud credits. AWS's self-funded Founders path begins at $1,000 and can reach $5,000; Portfolio applicants can reach $200,000 with an Activate Provider organisation ID. Google's AI programme offers up to $350,000 over two years to qualifying VC-funded AI startups and excludes third-party models. - [Microsoft for Startups](https://learn.microsoft.com/en-us/startups/microsoft-for-startups/getting-started-mfs): a milestone path up to $150,000 rather than one immediate grant. Microsoft's current application page says $200 at the start, while an older official activation page still says $1,000. Until Microsoft reconciles those pages, the honest claim is the staged maximum and the documentation conflict, not a precise opening cheque. - [DigitalOcean Hatch](https://www.digitalocean.com/blog/hatch-accelerate-your-ai-ml-startup), [OVHcloud Startup Program](https://startup.ovhcloud.com/en/startups/), and [Daytona Startup Grid](https://www.daytona.io/startups): selected-company programmes. DigitalOcean advertises up to $100,000 for its 12-month AI and ML programme. OVHcloud's public page currently supports EUR 10,000 for selected pre-seed or seed companies. Daytona advertises up to $50,000, but its linked terms currently return 404, so treat the headline as an application lead rather than bankable runway. - [Modal academic grants](https://modal.com/academics), [AMD's AI Developer Program](https://developer.amd.com/resources/), [Google research credits](https://cloud.google.com/edu/researchers), [AWS Cloud Credit for Research](https://aws.amazon.com/government-education/research-and-technical-computing/cloud-credit-for-research/), and [Hugging Face Community GPU Grants](https://huggingface.co/docs/hub/main/spaces-gpus): application routes for qualifying research and public work. Google's current public research route advertises $5,000. AWS caps student awards at $5,000 but not faculty or research-staff applications. Modal advertises up to $10,000. Awards, timing, regions, and permitted workloads remain programme-specific. ## Stacks that still fit together - **A small real server:** OCI for the VM, Cloudflare for DNS, CDN, and TLS, R2 or Tigris for objects, Neon or Aiven for Postgres, Resend for mail, and Grafana Cloud or Sentry for visibility. - **A no-server product:** Workers and static assets, D1 or Neon, R2, Queues or Workflows, Clerk or AuthKit, and Workers AI. The useful September change is that D1 now has predictable failure rather than a fuzzy edge after its daily allowance. - **An AI demo:** Paperspace or Colab for notebook work, Modal for repeatable burst jobs, ZeroGPU for a very small public Gradio surface, and Workers AI, Gemini, Groq, or OpenRouter for hosted inference. - **An agent's first deployment:** Railway or Cloudflare for the temporary application, Axiom for temporary logs, then Vercel Sandbox, Upstash Box, or Freestyle when the experiment needs repeatable code execution. Claim links are credentials in practice, and every unclaimed environment is designed to disappear. September did not kill a major category of free compute. It did something less dramatic and probably more useful: it put harder numbers around sandboxes, made one database's stopping behaviour explicit, and exposed several stale claims in the August list. The free tier is only useful when the boring corrections survive the monthly rewrite.