AWS vs Azure vs Google Cloud Free Trial Credits Compared
AWS vs Azure vs Google Cloud Free Trial Credits Compared
If you want frontier models on enterprise terms at no cost, the answer is rarely a lab's free tier. It is the cloud trial credit. AWS, Microsoft Azure and Google Cloud each front new accounts $200 to $300 of spendable credit that covers their hosted model APIs — Amazon Bedrock, Azure OpenAI, and Vertex AI respectively.
The headline amounts are remarkably similar. Everything else is not. The clocks differ by a factor of six, the model menus overlap only partly, the batch discounts change the effective value of the credit, and — most practically — each trial ends in a different way. This is the comparison a developer actually needs before signing up for one, two or all three.
How much free credit does each cloud give new accounts?
| AWS Free Tier | Azure free account | Google Cloud free trial | |
|---|---|---|---|
| Credit | Up to $200 ($100 at sign-up plus up to $100 earned through account activities) | $200, in your billing currency | $300 Welcome credit |
| Clock | Up to 6 months on the free plan | First 30 days | 90 days |
| Card required | Yes, for billing verification | Yes | Yes |
| AI surface | Amazon Bedrock, SageMaker, Amazon Polly | Azure OpenAI and the AI Foundry model catalogue | Vertex AI and the Gemini enterprise agent platform |
| Monthly free AI allowance | Polly Standard voices: 5 million characters per month during the free period | A limited set of free services for 12 months | 5,000 grounding queries per month, aggregated across Gemini 3 models |
| When the credit ends | The free plan does not continue on its own unless you move to a paid plan | Credit lapses; the account continues on normal billing | Credit lapses; standard pay-as-you-go continues |
Sources: AWS Free Tier, Azure free account charging guidance, Google Cloud free trial.
Three differences dominate the decision: the clock, the model menu, and what happens at expiry. Take them in order.
How we verified this
Every figure in this article was checked against the vendor's own pages on 2026-09-17: the AWS Free Tier page for the credit amount, the sign-up split and the six-month window; Microsoft's guidance on avoiding charges with an Azure free account for the $200 credit and the 30-day limit alongside 12-month free services; and Google's free trial page for the $300 Welcome credit and its 90-day validity. For AI-specific rates we used Amazon Bedrock pricing for the batch discount, Amazon Polly pricing for the character allowance, and Google's Agent Platform pricing for the grounding allowance. Where a vendor phrases an amount as "up to" or conditions it on eligibility, we have kept that phrasing rather than rounding it into a flat number.
The clock is the real price
The credit amounts are close enough to treat as equal. The windows are not, and they change the value of the credit entirely.
Azure's 30 days is a sprint. $200 of credit consumed inside a month is a lot of inference, but the constraint is time, not money. Divide it out and you have roughly $6.67 of runway per day. That figure forces a decision at sign-up: either schedule the heavy evaluation in week one and week two, or watch a material share of the credit expire unused. On Azure, an idle weekend is a real cost.
Google's 90 days is the middle path. The same shape of constraint, spread across a quarter, which happens to match a typical proof-of-concept or a quarterly evaluation cycle. Google states that you are not charged until you activate your full paid account, which makes the trial period relatively safe to leave running alongside other work.
AWS's six months is a runway. Here the money is the tighter constraint rather than the clock: up to $200 — $100 at sign-up plus up to $100 earned through account activities — spread across half a year. If you are building a real project rather than running a one-off evaluation, this is the trial whose clock matches a project timeline rather than an evaluation sprint.
The rule that follows: match the trial's clock to your workload's duration, not to its headline number.
| Workload duration | Best-fit trial | Why |
|---|---|---|
| Two-week evaluation | Azure | Fastest path to the GPT model line, and 30 days is ample |
| One quarter of RAG development | Google Cloud | 90 days matches the cycle, plus a recurring monthly grounding allowance |
| Six-month side project | AWS | The only trial whose clock spans the project |
| Ongoing, low-volume experimentation | AWS plus Google Cloud | The long AWS clock carries steady work; Google's monthly grounding allowance is recurring |
Set your decision date at half the window
The most common way to lose a trial credit is not overspending. It is discovering the expiry the week the bill starts. Set a calendar reminder at 50% of each window — day 15 for Azure, day 45 for Google, month 3 for AWS — and treat that date as a decision point with three options: convert to a paid plan, migrate the workload to a first-party free tier, or shut the environment down cleanly.
Add a second reminder at 80% of the window for the same reason. Budget alerts help too, but they fire on spend, not on elapsed time, and a trial that lapses unused triggers no alert at all.
What the credit actually buys for AI work
The credit is spendable across each cloud's broader AI surface, which makes the same $200 buy three different model menus.
AWS / Amazon Bedrock. A multi-vendor catalogue served through one API, alongside SageMaker and Amazon Polly. Two features change the arithmetic. First, Bedrock's batch inference is documented at 50% of on-demand pricing, so credit spent on offline evaluation covers roughly twice the inference volume (Bedrock pricing). Second, Polly's free tier includes 5 million characters per month for Standard voices, which is a standing allowance during the free period rather than part of the credit (Polly pricing).
Azure / Azure OpenAI. Access to the GPT model line through Azure deployments, with residency and data-processing terms that come from Azure rather than from the model vendor. Alongside the 30-day credit, Microsoft's documentation describes a limited quantity of free services for 12 months. If your requirement is specifically "OpenAI models under an enterprise cloud agreement", this is the direct route.
Google Cloud / Vertex AI. Gemini plus partner models under Google Cloud terms, and the piece none of the others match: 5,000 grounding queries per month at no charge, aggregated across Gemini 3 models, with additional queries billed per thousand (Agent Platform pricing). That allowance is recurring and independent of the trial credit, which makes it the most durable free line item in this comparison.
Worked example: how far does $200 of batch credit go?
Batch pricing is the single biggest multiplier on trial credit, and the arithmetic is simple. If batch inference runs at 50% of on-demand pricing, every dollar of credit buys two dollars of on-demand-equivalent inference.
| Credit available | On-demand-equivalent inference | If spent on batch |
|---|---|---|
| $200 (AWS or Azure) | $200 | $400 |
| $300 (Google Cloud) | $300 | $600 |
Now make it concrete. Suppose your evaluation plan has an on-demand cost of $600 — a full re-run of a benchmark across a document corpus, for example. Run it as a batch job and it costs $300. That is a third more than Google's $300 credit covers in full, and half again more than AWS's $200 credit can absorb; run the same plan on-demand and the $200 credit covers just a third of it. The same credit, spent on batch, is worth twice as much evaluation.
The grounding allowance deserves its own arithmetic because it is the item most likely to surprise you. Suppose your RAG evaluation issues 12,000 grounding queries in a month:
- First 5,000 queries: covered by the monthly allowance at no charge
- Remaining 7,000 queries: 7 × $14 = $98 for that month
- Repeating that evaluation twice inside a 90-day window: $196
Two months of a moderately sized grounded evaluation consumes roughly two thirds of Google's entire $300 credit. That is not an argument against the trial — it is an argument for measuring your grounding volume before you scale it. Google also warns that a single user request can expand into more than one search query and that you are billed per individual query, so query volume can exceed prompt volume.
What happens when the credit runs out
This is where the three trials diverge most sharply, and where a forgotten account can create real trouble.
- AWS: the free plan is explicitly time-limited. AWS describes it as up to six months for new customers with up to $200 in credits, and the account does not continue on its own afterwards unless you move it to a paid plan. Check the AWS Free Tier FAQs for the current closure behaviour and migrate anything you intend to keep before the window closes.
- Azure: the credit ends, the account does not. Microsoft's guidance states the credit is for the first 30 days, so unused credit is forfeited at that point while the subscription continues on normal billing for anything still running.
- Google Cloud: the credit lapses and standard pay-as-you-go continues. The project and its resources survive; only the free money disappears.
The practical asymmetry: AWS is the trial most likely to close on you, and Azure is the trial most likely to catch you unprepared, because 30 days is short enough to forget.
A playbook for running more than one trial
Many developers qualify for all three. Doing it deliberately rather than accidentally looks like this:
- Keep the trials isolated. One cloud per evaluation, one billing profile each, no production dependency on any of them.
- Write the three expiry dates into a single calendar entry, with reminders at 50% and 80% of each window.
- Enable budget alerts immediately, before the first workload. Alerts are not a substitute for the date reminders; they cover different failure modes.
- Decide at the halfway point. Convert, migrate or terminate — but decide, rather than letting the clock decide for you.
- Move durable workloads to first-party free tiers once the evaluation concludes, so a trial expiry cannot take a live product offline.
- Verify eligibility per vendor. "Up to" amounts and new-customer definitions are vendor-specific, and one trial does not disqualify you from another vendor's.
Limitations to keep in mind
- Trial amounts are conditional. "Up to $200" includes an earned portion on AWS, and eligibility differs by vendor and by customer history.
- Eligibility is per customer, not per project. One trial per eligible customer means you cannot spin up a second to extend the clock.
- Card verification is standard. All three require billing verification, and the trial's protection is the credit, not the absence of a payment method.
- Some services are excluded from the credit, and regional availability varies — particularly for the newest models in each catalogue.
- Always-free tiers are separate from trial credit. A service with a permanent free allowance behaves differently from one funded by a depleting balance; do not assume an always-free line item is protected by the trial.
- Model catalogues and prices change. Verify the specific model, region and mode you intend to use on the vendor's own pricing page on the day you commit, because none of these numbers is stable across quarters.
FAQ
Can I extend a cloud trial credit? Not by asking. The windows are fixed for eligible new customers. The practical alternatives are a startup credit programme, which each of the big three runs with application-based eligibility, or stacking a first-party free tier from a model lab for the post-trial period.
Which trial is best for evaluating OpenAI models specifically? Azure, because its credit is directly spendable on Azure OpenAI deployments under Azure's terms, and 30 days is enough for a focused evaluation if you schedule it deliberately.
Which is best for a RAG prototype? Vertex AI. The 90-day window suits a quarter of development, and the recurring 5,000 grounding queries per month is a free retrieval allowance no other trial includes.
Does using batch inference delay my results? Yes, and that is the trade. Batch work is asynchronous by design, so it suits evaluation, backfills and bulk re-processing rather than interactive features. In exchange, the same credit covers roughly twice the volume.
Do I need to cancel the account when the trial ends? Not necessarily, but you must decide. On Azure and Google Cloud, leaving the account running starts normal billing for anything active. On AWS, the free plan does not continue indefinitely, so migration is a prerequisite rather than an alternative.
Where this sits in the directory
A directory entry that reads "AWS: free" hides the six-month plan boundary. "Azure: free" hides the 30-day clock. "Google Cloud: free" hides that the credit is time-boxed and that the durable free value is the recurring grounding allowance, not the headline number.
That is why the Amazon Bedrock entry, the Azure OpenAI entry and the Vertex AI entry each carry the exact credit amount, the validity clock and the expiry behaviour in their free-tier notes. On enterprise cloud terms, the trial credit is only as good as the calendar you set for it — so set the calendar before you set the model.
- How much free credit does each cloud give a new account?
- Google Cloud's free trial preloads $300 in Welcome credit valid for 90 days. AWS offers new customers up to $200 in credits across a six-month free plan. Azure's free account gives eligible new customers $200 in credit for the first 30 days, alongside a limited set of free services for 12 months.
- Which cloud trial lasts the longest?
- AWS's free plan runs up to six months, Google's trial credit is valid for 90 days, and Azure's credit expires after 30 days. Match the clock to your project: a two-week evaluation fits any of them, a three-month prototype needs Google, and a half-year project needs AWS.
- Can I claim all three trial credits?
- They are separate vendors with separate account and eligibility rules, so a person or organisation that qualifies for each can hold all three. The real cost is operational: three billing surfaces to monitor and three expiry dates to action before charges begin.
- What happens to my resources when the trial credit runs out?
- On AWS the free plan is time-limited and does not continue on its own unless you move to a paid plan. On Azure and Google Cloud the credit lapses and normal pay-as-you-go billing takes over, so the account and its resources survive but start costing money.
- Does batch inference make trial credit go further?
- Yes, substantially. AWS lists batch inference on Amazon Bedrock at 50% of on-demand pricing, and Google lists a 50% batch reduction on its Gemini API pricing, so a credit spent on batch work covers roughly twice the inference volume it would otherwise.
- Are trial credits enough to evaluate frontier models properly?
- Usually yes for an evaluation and often yes for a small prototype, but not for sustained production traffic. Size the evaluation before you start: a workload running a few million tokens a day can consume most of a $300 credit inside 90 days, and the clock runs whether or not you are testing.