What free AI credits are actually worth

Sooner or later someone offers you free AI. It arrives as a startup credit, a research grant, a nonprofit discount, an accelerator perk, or a founder in a group chat saying his company got $200,000 of cloud for filling in a form. The number is always large, the form is always short, and the offer is always genuine. What it is not is free, and the part that costs you is never in the headline figure.

This guide is about how to take that money well. Not whether to take it, because usually you should, but how to price it in your own work, which four terms in the fine print decide what it is worth, and how to arrive at the day it ends without a rebuild on your hands. It is written for a solo operator or a small team who would notice a $600 monthly bill appearing. If you have a procurement function negotiating committed spend with a vendor’s sales team, your version of this problem is a contract negotiation and this is too small for you.

A credit grant is a marketing budget with a meter attached

Read the offers on their own terms first. Anthropic’s AI for Science rare disease grants gave researchers up to $50,000 in Claude credits over six months, split into a basic-science track built around the Monarch Initiative’s disease ontology and knowledge graph, and a biotech track aimed at drug development, with named existing grantees and an application deadline of August 2, 2026 at 11:59 PM PST [1]. That round is closed, which is itself the first thing to learn: these programs run in rounds, and the one you read about is usually the one you have just missed. The program also let qualifying projects request exemptions from the bio classifiers, an admission that safety systems built for adversarial use will trip on legitimate biology [1].

Now read the same offer as a budget line. The cloud versions are much larger. Google for Startups Cloud Program lists up to $200,000 in credits at its Start tier and up to $350,000 at Scale [5]. AWS Activate lists up to $200,000 [6]. Microsoft for Startups lists up to $150,000 across eligible Azure services [7]. No company gives away six figures out of kindness alone. What it buys is a workload that starts on its platform during the months when you are deciding how things will work, plus real usage data, plus, if the work goes well, a case study that a competitor’s marketing team cannot manufacture because the underlying relationship is real.

Both readings are true at once and neither cancels the other. The rare-disease researcher gets compute she could not otherwise afford, and the lab gets evidence its model does hard scientific work. Your job is not to pick a side. It is to notice that you are being paid to make a decision early, and to make sure the decision you make is one you would still make at list price.

Four terms decide what a grant is worth

Everything that matters is in four questions, and vendors answer all four in public documentation.

The first is when the credits die. Anthropic’s ordinary prepaid API credits expire one year from the purchase date, the expiration cannot be extended, and purchases are non-refundable [4]. Grant credits are usually shorter: the rare-disease grants ran over six months [1]. A window shorter than your project means the unspent balance is not savings, it is a number that goes to zero on a date. Write that date somewhere you will see it.

The second is what the credits cover. Google’s program states plainly that credits apply to Google Cloud services and cannot be applied to third-party services or offerings, including those on Google Cloud Marketplace, and that third-party models are billed directly and are not covered [5]. AWS goes the other way and makes its credits redeemable on third-party models available through Amazon Bedrock [6]. That single difference decides whether a grant funds the architecture you actually wanted or quietly steers you into the one the sponsor prefers.

The third is what the credits are denominated in, which is always list price. A grant of $50,000 in credits is $50,000 of that vendor’s rate card on the day you spend it, not $50,000 of compute in the abstract. On current Claude API pricing, Opus 5 is $5 per million input tokens and $25 per million output, Sonnet 5 is $2 and $10, and Haiku 4.5 is $1 and $5, with batch processing at half the standard rates [3]. Spend the grant on the largest model because it is free and you burn it four to five times faster than the same work on a mid-tier model, and you learn nothing about whether the cheap one would have been fine.

The fourth is what happens the morning after. Claude for Startups is direct about it: standard API pricing kicks in automatically with no migration [2]. There is no cliff in the engineering sense, which is exactly the problem. Nothing breaks, nothing prompts you, and the bill simply starts. On the consumer side the equivalent hard stop is more honest: run out of prepaid API credits and you cannot call the API at all until you add more [4].

Price the grant in your own workload, not in dollars

The headline number tells you nothing until you divide it by the way you actually work. Do the division in the direction that hurts, which is not “how many months will this last” but “what is my bill the month after it stops”.

If a grant is $10,000 over 6 months, you have been handed a run-rate of about $1,667 a month. Some of that is work you would never have done at your own expense, which is fine and is part of what a grant is for. The rest is the load-bearing part: the daily classification job, the support drafting, the thing a client now expects. Estimate honestly what share of the usage you would keep paying for, multiply, and look at the answer before you build anything.

If that number is comfortable, take the grant and stop worrying. If it is not, you have found the real constraint early, and you have several months of subsidised time to fix it: move the routine work to a cheaper model, batch what does not need an answer in the next few seconds at half price [3], or discover that a free tier covers it. Gemini’s free tier covers its 3.x models with limited access, and paid rates on the cheaper tiers start well below the frontier ones, with Gemini 3.5 Flash-Lite at $0.30 per million input tokens and $2.50 output [8]. Plenty of small operations never leave that band.

Spend the credits on work, not on habits

There is a specific failure that follows free credits around, and it is not overspending. It is that you stop making decisions. Retrieval gets sloppy because context is free. Every request goes to the biggest model because why not. A nightly job reprocesses everything instead of the delta. None of it hurts while the meter is someone else’s, and all of it becomes your run-rate on a Tuesday in month seven.

The fix is to run the grant period as if you were paying, and let the credits buy you volume rather than carelessness. Keep the console’s usage view open often enough to know your per-day burn, because that number, not the total, is what you will be renewing. Do the experiments you could not have afforded, which is the genuine value here: run the same task across three models and keep the transcripts, try the expensive one on the hard 5 percent and the cheap one on the rest, process the backlog you have been putting off. Those are one-time spends that leave you with knowledge and a finished backlog rather than a habit.

Write down, in a file, what each recurring job costs per run at list price. It takes an hour and it is the single thing that makes the end of the grant boring instead of alarming.

Build so the credits are fuel and not foundation

Lock-in from a credit program is rarely contractual. It is architectural, and it accumulates quietly.

Keep prompts and evaluation cases as text files in your own repository rather than as objects inside a vendor’s console. That one habit is most of the portability you need, because prompts are the part you actually spent time on and they move between Claude, ChatGPT and Gemini with edits rather than a rewrite. Route model calls through one small function of your own instead of scattering the vendor’s client across a dozen files, so that switching is a change in one place. Keep a set of 20 to 50 real inputs with the outputs you consider correct; without them you cannot tell whether a cheaper model is good enough, and you will default to the expensive one out of uncertainty.

Then check the things that are genuinely hard to move, because those are where the sponsor’s money is really aimed. Data that lives in the sponsor’s storage moves slowly and sometimes expensively. A vector index built in their service is a rebuild elsewhere. Fine-tuned or otherwise adapted models generally do not travel at all. Using a credit grant on stateless inference is close to free in switching terms. Using it to build your data layer inside one vendor’s platform is a decision you are making for years, on a subsidy that lasts months.

Where the programs are and whether you qualify

Four categories cover almost everything on offer, and the eligibility rules are narrower than the marketing suggests.

Startup credits are the largest and the most conditional. Claude for Startups is open to any early-stage founder building with Claude, but credits require that the company received equity funding from an institutional investor, was founded within the last 4 years, and has not previously received Anthropic startup credits, with priority rate limits as the other benefit and an application described as taking about 2 minutes [2]. The cloud programs are the six-figure ones [5][6][7]. If you are bootstrapped and unfunded, most of this tier is closed to you, and it is worth knowing that before you spend a morning on applications.

Research and science programs are narrower in scope and more generous in spirit. The rare-disease grants required a specific kind of work, named their partners and grantees, and set a hard deadline [1]. Programs of that shape reward a real project description and punish a vague one. Nonprofit and education pricing exists at most vendors and usually takes the form of a discount rather than credits, which is better, because a discount does not expire. And free tiers are the category people skip because they feel like a lesser thing; Gemini’s free tier is available across its current models [8], and for a solo operator running a few hundred tasks a month, the honest answer is often that you did not need the grant.

One more use for these programs, whether or not you qualify: they are a vendor-vetting signal. A lab willing to name its partners, its existing grantees, a specific dollar figure and a real deadline is exposing itself to being judged on outcomes [1]. That tells you more than a benchmark score does, because outside researchers are already using the tool on hard problems and the lab is comfortable being publicly attached to the result.

What still goes wrong

Every figure here is a snapshot verified on the date in the header, and credit programs change faster than pricing pages do. Rounds open and close, amounts move, eligibility gets tightened when a program is oversubscribed, and a page that promised up to $350,000 in March can promise something else by autumn. The rare-disease round described above has already closed [1]. Check the source before you plan around any of it, and treat a program you read about in an article, including this one, as a starting point for a search rather than a live offer.

The arithmetic also has a blind spot it cannot fix. It tells you what the work will cost, not what it will be worth, and those are different questions. A subsidised six months that produces one system your clients now depend on is a good trade even if the ongoing bill is uncomfortable, and a subsidised six months that produces impressive demos is a bad trade at any price. Only you can tell which one you had, and you can usually only tell afterwards.

Finally, a caution about the caution. This guide optimises for reversibility, which is the right bias when you are spending your own money and a bad quarter matters. It is the wrong bias if the entire point is to build something deep on one platform, where committing early and hard is the strategy rather than the risk. Frontier labs offer these programs precisely because a fraction of recipients will commit, and some of them are right to. Just make the commitment deliberately, on the day you can still see the list price, rather than discovering in month seven that you made it without noticing.