The Current Credits Index: AI compute free credits, August 2026

Live version: the credits index. This page is a frozen snapshot, checked 2026-08-21; the index behind that link is the version that gets re-checked and re-dated. Snapshots are taken monthly, one dated URL per month, so this report is never edited in place once published.

Update 2026-08-24: UpCloud is currently advertising a $500 / €500, 30-day promotional free trial; the live UpCloud entry tracks it. Everything below, including the $300-topped grant ranking, reflects the original 2026-08-21 snapshot and is intentionally left unchanged.

Short versionCredits index

Rent	When	Example
Agent sandboxes	The largest card-free grants live here: Daytona $200 (CPU sandboxes only), E2B $100, Novita $100 (Sandbox scope, 90 days).	Daytona
GPU training	No provider publishes a free GPU-VM pot. The real offer is recurring serverless credit: Modal and Beam, $30 per month each.	Modal
API tinkering	Fireworks gives $1 that spends on inference. OpenRouter, Groq, Mistral, Cohere, and Cloudflare publish $0 paths with limits.	Fireworks AI
VPS self-hosting	Two $300 grants (Google Cloud, Oracle) and a $200 band (AWS, Azure, IBM, OVHcloud). Oracle Always Free stays $0 after the trial.	Oracle Cloud

38 of 55 checked providers publish something real. 17 publish no free credit at all.Checked 2026-08-21

This is the Current Credits Index for August 2026: a dated, citeable snapshot of every AI compute free credit, free tier, and trial this site has verified against official pages, plus every circulated credit figure that those pages do not back. The living version is the credits index, where each entry is re-checked and re-dated on its own schedule. This report freezes the state of the landscape as checked on 2026-08-21 so it can be linked, quoted, and compared against later snapshots.

Findings

I checked 55 providers. 52 entries were verified on 2026-08-21; three (DigitalOcean, E2B, Modal) carry a lastChecked of 2026-08-20. Of the 55:

• 38 publish something real: 14 one-time signup credits, 9 monthly free tiers, 8 time-boxed trials, and 7 always-free tiers.
• 17 publish no free credit at all. Most of those entries exist because the opposite claim circulates: a specific dollar figure attached to a provider whose official pages state no such thing.
• 21 circulated figures fail verification and are corrected in the table below, including the widely reprinted OpenAI $5, Anthropic $5, RunPod $10, Replicate $5 or $10, and Crusoe $300.

The biggest verified grants are two $300 offers: the Google Cloud Welcome credit, $300 / 90 days, and the Oracle Cloud trial credit, US$300 over up to 30 days. Behind them sit Civo at $250 for 31 calendar days and a $200 band: AWS at up to $200 over a 6 month free plan, Azure at $200 for 30 days, IBM Cloud at USD 200 after a Pay-as-you-go upgrade, OVHcloud at US$200 for one month, and Daytona at $200, no card for CPU sandboxes.

The most useful grants are not always the biggest. For anyone unwilling to enter a payment method, the ranking changes completely: Daytona $200, E2B $100, and Novita AI $100 are the three largest card-free grants, and all three are sandbox products. For sustained work, the only recurring dollar credits in the set are Modal and Beam, each at $30 per month of serverless compute.

One structural finding deserves the headline: there is no free GPU-VM money anywhere in this set. Every path to a GPU either requires your own deposit (RunPod, Vast.ai, TensorDock), excludes GPUs from the free grant (Daytona, AWS free plan, Oracle Always Free, UpCloud trial), zeroes the GPU quota during the trial (Azure, Google Cloud), or scopes the credit to serverless endpoints (Modal, Beam, Crusoe).

Method

What counts as a source

Official first-party pages only: pricing pages, documentation, legal terms, and official signup surfaces. Blog roundups, coupon aggregators, YC threads, and Reddit posts are treated as claims to verify, never as sources. Every entry in the credits index lists the official URLs it rests on, and the sources list on this page carries the official URLs behind this report’s headline claims. Where a provider’s official pages could not be reached or do not state a number, that is recorded as a finding, not papered over.

Fail-closed rules

The index is fail-closed. The rules, applied uniformly:

1. If an official page does not state a number, the published value is “Not stated officially”. That string appears throughout the tables below and is a legitimate cell value, not a gap.
2. A minimum deposit is not a credit. “Start with as little as $10” (RunPod) and “The minimum deposit amount on Vast.ai is $5” describe your money, not the provider’s.
3. A spend cap is not a credit. OpenAI’s usage-tier table includes a tier named Free with a $100 per month usage limit; that is a ceiling on permitted usage, not $100 of free money.
4. A $0 plan fee is not a credit. Cerebrium’s Hobby plan is “Free + compute / month”: the plan costs nothing, every compute second bills.
5. Application-gated programs are not signup offers. AWS Activate, Microsoft for Startups, Lambda’s research grants, and the various startup programs are recorded as context, never as the default grant.
6. Nothing here is a usage report. Every entry’s hands-on field reads “Not tested yet” until that changes; this report fabricates no experience.

Check dates

All 55 entries were verified within the two days ending 2026-08-21: 52 on 2026-08-21 itself, and three (DigitalOcean, E2B, Modal) on 2026-08-20. Each linked entry displays its own lastChecked date, and that per-entry date is the authoritative one if this snapshot and an entry ever disagree.

What was excluded and why

Four providers were checked but have no published entry, so they are excluded from every count above:

• Baseten. Official docs say “New workspaces receive credits for testing and deployment” but publish no dollar amount. A credit that exists with an unpublishable size fails rule 1.
• Lightning AI. Official docs say eligible free plan users receive complimentary credits with no credit card required, but the dollar figure is not on a crawlable official page; the pricing page renders as a JavaScript shell. Excluded until the number can be read from an official source.
• Google AI Studio / Gemini API. The official rate-limit, pricing, and billing pages redirected to a Google sign-in during checking. Third-party numbers for 2026 conflict with each other, so nothing is published.
• Genesis Cloud. The marketing site returned 522 and 404 errors and the console was down on 2026-08-21. There is no live official page to verify anything against.

A further four GPU clouds (Hyperstack, CUDO Compute, GMI Cloud, and Verda, formerly DataCrunch) were checked and found to publish no automatic signup grant, but their entries are not yet published, so they sit outside this snapshot’s 55.

When a provider disagrees with itself

Where official pages state the same fact twice differently, both versions get printed and the dated documentation page wins over the marketing page:

• Google Cloud describes the Welcome credit window as 90 days on the program pages while the legal terms say 3 months. The entry prints both.
• Civo’s terms grant the $250 for “31 calendar days” while marketing copy says “a full month”. The terms figure is published.
• Cerebrium’s pricing page states 1 day of log retention in one place and 7 days in another. The entry flags the contradiction rather than picking silently.
• DigitalOcean’s billing docs describe per-second billing with a 60 second minimum while a marketing page still says 5 minutes. The docs, updated 2026-08-18, win.
• AWS advertises “up to $200 in credits”, which official pages themselves decompose into $100 at signup plus up to $100 earned by exploring key services. A new account that does nothing has $100, not $200.