Can open-source maintainers get free AI API credits?

Yes. Several vendors now run dedicated open-source programs that are separate from their public free tiers. They split into two kinds: credits for the maintainer’s own workflow — code review, triage, release automation — and credits for the inference your project runs for its users. Both are applied for rather than signed up for: a person reads the application, the amount is sized to the project, and the grant is time-bound.

What is the difference between a free API tier and an open-source credit grant?

A free tier is self-serve, identical for everyone, sized for evaluation, and needs no approval. A grant is an application a person reviews, sized to your project’s reach and to what you say you will build, usually time-bound, and usually carrying an obligation — visible attribution in your README, or using the vendor’s tool in your maintainer workflow. A free tier lets you try a model; a grant lets you ship a feature.

Do I need GitHub stars to qualify for AI credits?

Not necessarily. Some programs band their tiers by stars; others never mention stars and measure reach by dependent repositories, monthly downloads across a package registry, merged pull requests, unique external contributors, or an OpenSSF criticality score. Where stars are used they set a starting tier rather than the decision, and at least one program publishes an explicit tier for projects under 100 stars — so a small but genuinely depended-on project is not automatically out.

Can I apply to more than one AI credit program at the same time?

Generally yes, and for many maintainers it is the right move, because workflow credits and product-inference credits pay for different things. Read each program’s own terms first — some require you to use their tool in your maintainer workflow, or to keep an attribution badge in your README, as a condition of the grant. Apply to each with a concrete, program-specific answer rather than one paragraph copied around.

AI Credits for Open Source Maintainers: 5 Programs Compared

AI credits for open source maintainers went from a rumour to a category in about a year. Five vendors now run a program you can apply to, each with its own bar and its own idea of what the money is for — and from the outside they look interchangeable. They are not. Three of them fund the AI you use to maintain the project; two fund the AI your users trigger inside it. Nobody had put the five side by side, so here they are, every cell checked against the program’s own page. One of the five is the ModelRunner Open Source Program, which we run — it gets one row like everything else.

The AI Credit Programs Open-Source Maintainers Can Apply To in 2026

Program	What you get	Who qualifies	Term	What it funds
OpenAI — Codex for Open Source	Six months of ChatGPT Pro with Codex, conditional access to Codex Security for core maintainers with write access, and API credits through the Codex Open Source Fund.	Core maintainers, or anyone running a widely used public project. Projects that do not fit the criteria are invited to apply anyway and explain why.	Six months on the ChatGPT Pro seat. No term is published for the credits.	Your maintainer workflow — the credits are scoped to Codex in pull request review, maintainer automation and release workflows.
Anthropic — Claude for Open Source	Six months of Claude Max 20x, free.	Five published routes, any one of which qualifies: dependent-repo, dependent-package or monthly-download thresholds; being a listed committer on a recognised foundation or language project; 100+ pull requests merged into repositories you do not own in the last 12 months; a repository with 20+ unique external contributors in the last 12 months; or an OpenSSF criticality score of 0.4 or above.	Six months, then the complimentary subscription ends. A prior paid plan resumes at its old rate; otherwise the account returns to the free plan.	Your maintainer workflow — it is a subscription seat, not API credit.
Ona for Open Source	Up to $200 a month in credits.	Maintainers and core contributors of at least one active, well-established open-source project. The amount granted depends on the degree of your open-source contributions.	Monthly.	Your maintainer workflow — the program is framed around managing AI-generated pull requests, clearing a backlog and dependency updates.
Atlas Cloud for Open Source	$50–150/mo at 100–1,000 stars, $150–400/mo at 1,000–10,000 stars or 100k+ monthly downloads, $400–1,500+/mo at 10,000+ stars or ecosystem-critical infrastructure. Plus referral earnings and co-marketing.	Primary or core maintainers of public, actively developed projects under an OSI-approved licence, updated within the last 3 months. One application per project.	Starter and Growth are one-time 3-month grants; Flagship renews monthly.	Inference inside your project — image, video, audio, 3D and language models through one key.
ModelRunner Open Source Program	$25 under 100 stars, $50–150 at 100–1,000 stars, $150–400 at 1,000–10,000 stars or 100k+ monthly downloads, $400–1,500+ at 10,000+ stars or ecosystem-critical infrastructure — every tier a one-time credit.	Primary or core maintainer; public repository under an OSI-approved licence; a commit within the last 3 months; one application per project, from the email on your account.	Every tier is a one-time credit; nothing renews automatically, and you can re-apply as the project grows.	Inference inside your project — image, video, audio, 3D and language models through one API.

Every cell is taken from the program’s own page and was checked on 7 September 2026. Programs change their terms without notice — read the program’s page before you apply. Rows are grouped by what the credits fund, not ranked by size.

Four things worth reading out of that table:

Two different products share one label. Three rows hand you a subscription seat or an allowance for an agent that works on your repository. Two hand you API credit your own code spends when a user hits a feature. Calling both “AI credits for open source” hides the only distinction that changes what you can build.
Reach is measured four incompatible ways. Stars, dependent repositories and registry downloads, merged pull requests to repositories you do not own, and an OpenSSF criticality score all appear as published bars. A project that reads as “too small” against a star band can clear a dependents threshold comfortably, so check the programs individually rather than in one pass.
Delivery mechanics differ, and decide how the money reaches you. A complimentary subscription applied to an existing account, a monthly credit allowance, a redemption code emailed after approval, or balance added to the account you already bill from — each implies a different amount of plumbing before a single call runs.
Nothing here is permanent funding. The published terms are six months, three months, a monthly renewal or a single one-time credit. A grant buys a runway to ship something and see whether it gets used; it is not a budget line you can plan years against.

Free Tier vs Open-Source Credit Grant: Why They Are Not Interchangeable

Almost every AI vendor also publishes a free tier, so maintainers reasonably ask why a grant is worth an application. They are different instruments. A free tier is self-serve and identical for everyone: no review, no negotiation, sized so you can evaluate a model and decide whether to pay for it — deliberately too small to run a feature on.

A grant is an application. A person reads it, sizes the amount to your project’s reach and to what you said you would build, and attaches a term and usually an obligation — a badge in the README, or using the vendor’s tool in your workflow. The trade is real scrutiny for real headroom: a free tier lets you try a model, a grant lets you ship a feature other people use.

Which Program Fits What You Are Building

This is the question the five landing pages cannot answer for you, because each of them only knows about itself. Sort the programs by whose usage the credits absorb — yours, or your users’.

Credits That Fund Your Maintainer Workflow

OpenAI’s Codex for Open Source, Anthropic’s Claude for Open Source and Ona for Open Source all pay for the AI you run while maintaining the project: reviewing incoming pull requests, triaging issues, keeping dependencies moving, automating a release. Consumption scales with how much maintenance you do, so a busy repository with no product surface at all is a perfectly good fit.

Read the conditions, because for two of the three the tool is part of the deal rather than a detail. OpenAI’s page scopes its API credits to projects that use Codex in pull request review, maintainer automation and release workflows; Ona’s program exists to put its own agent between contributors and your merge queue. That is not a catch — it is what the program is buying — but the grant arrives with a workflow attached, and it is worth wanting that workflow before you apply.

Credits That Fund Inference Inside Your Project

Atlas Cloud and ModelRunner pay a different bill: the calls your users trigger. If your project generates an image, transcribes audio, renders a 3D asset or answers a question with a language model, that usage scales with your users rather than your commit rate, and a maintainer-workflow seat does nothing for it. These are the programs that let a popular tool turn on an AI feature without its author personally underwriting the traffic.

The sizing question flips accordingly. For a workflow grant you are estimating your own week. For an inference grant you are estimating how often your users hit the feature and what one call costs — arithmetic you can do in advance, because per-model rates are published.

What Reviewers Check Before Approving a Grant

All five are reviewed by a person, and all five publish enough about their criteria to work out what that person is looking at.

The Four Signals Reviewers Screen For

A licence. Two of the five state an OSI-approved licence outright as a condition; the other three describe the project instead. A public repository with no licence file is not open source in the sense any of them mean, so fix that before you apply anywhere.
Recent activity. The published windows are short and explicit: a commit within the last 3 months on two pages, contribution activity within the last 12 months on another. A dormant repository with a large historical star count is what they filter out.
Your role. Most ask for a primary or core maintainer, or a listed committer. One is deliberately wider and accepts core contributors too, sizing the grant to how much you contribute — look there first if your work lands in repositories you do not own.
Reach, in whichever unit that program publishes. Stars, dependent repositories, dependent packages, monthly downloads, merged pull requests, unique external contributors, an OpenSSF criticality score. Where stars are used they set a starting tier, not the decision — usage, dependents and ecosystem impact move it.

Writing the “What Will You Build” Answer

Every application has a free-text box, and it decides the outcome once you clear the bar. ModelRunner’s form asks for “the AI feature you want to ship, the models or modalities you need, and roughly how your users would hit it”; Atlas Cloud asks for your repository, your role and how you plan to put the credits to work. Same question. Answer it in three parts.

Name the feature, not the ambition
“Generate a preview thumbnail for every template in the gallery” is reviewable. “Explore AI capabilities for our users” is not. A reviewer is deciding whether the credits turn into something that ships.
Name the modality and roughly what it costs
Say whether it is image, video, audio, 3D or a language model, and which end of the price range you expect. It makes a requested amount look considered rather than guessed.
Say how often your users would hit it
Downloads per month, active installs, requests a day — any real number. It turns your reach into a usage estimate, which is what the reviewer sizes the grant against.

The README Badge Convention

The two programs that fund inference inside your project both ask for a “Powered by” badge in the README while the grant is active, and on ModelRunner’s side removing it pauses the grant. The three maintainer-workflow programs do not ask for one.

The split is not arbitrary. A badge tells a project’s users where its AI features actually run, and it is how most maintainers find these programs in the first place — through somebody else’s README rather than through search. A workflow grant produces nothing user-visible to attribute, so there is nothing to badge. Treat it as the disclosed price of the inference programs, not as fine print.

When the Grant Ends: What Each Program Leaves You With

Every one of these grants ends, and the five endings differ more than the headline amounts do — which matters if a feature is about to depend on one.

The two subscription programs expire on a date. Six months, after which the ChatGPT Pro or Claude Max seat stops. Anthropic’s page is specific about the landing: you get an email first, a prior paid plan resumes at its old rate, and otherwise the account drops to the free plan.
The monthly programs continue while you qualify. Ona publishes its credits as a monthly amount, and Atlas Cloud renews its top tier monthly while making the middle tiers a one-time three-month grant. ModelRunner grants a single credit at every tier, which you re-apply for as the project grows.
The inference programs leave a working account behind. When ModelRunner credits are spent the account keeps working on ordinary usage-based billing at the published per-model rates — the same rates every account pays, and the same rates the credits were spent at, so nothing about your cost per call changes on the day the grant ends. Credits themselves expire 365 days after they are added.

Sizing a grant is arithmetic rather than a guess: multiply the per-call rate for the model you chose by the traffic you expect. Our breakdown of what a generated image actually costs runs that calculation across fifteen models, and the same method sizes any grant on this page before you commit a feature to it.

Applying to the ModelRunner Open Source Program

Our row, stated as plainly as the other four. Credits are granted against the four tiers below, decided by human review.

Tier	Criteria	Amount	Term
Seed	Under 100 stars, project at least 3 months old	$25	One-time credit
Starter	100 – 1,000 GitHub stars	$50 – $150	One-time credit
Growth	1,000 – 10,000 stars, or 100k+ monthly downloads	$150 – $400	One-time credit
Flagship	10,000+ stars, or ecosystem-critical infrastructure	$400 – $1,500+	One-time credit

Tiers as published on the ModelRunner Open Source Program page. Stars set a starting tier; actual usage, dependents and ecosystem impact decide the final one.

Eligibility is four lines: you are a primary or core maintainer, the project is public and under an OSI-approved licence, it has a commit within the last 3 months, and one application per project comes from the email on your account. Approved credits are added to that account’s balance — no separate credit currency, no coupon code — so requests draw it down at each model’s published rate. The catalog covers image, video, audio, 3D and language models through one API, over REST, the JavaScript SDK, or the hosted MCP server. A person reads every application and usually replies within 3–5 business days.

One row is worth naming, because it is a checkable difference rather than a claim: Seed is the only tier in this comparison published for projects under 100 stars. Keep it in proportion — three of the five programs do not band by stars at all, so a small project is not automatically out of those either.

If the bill you need covered is the one your users generate, apply to the ModelRunner Open Source Program with your repository, your role and the feature you want to ship. If it is the bill your own maintenance work generates, one of the other three rows is the better application — a fine outcome for a table we published.

How this comparison was checked

Every figure, eligibility line and term was read from the operating vendor’s own page on 7 September 2026, never from a third-party roundup. Several figures in wide circulation do not appear on the first-party pages at all, and inheriting a directory’s numbers would inherit its staleness.

A claim that could not be verified first-hand was dropped rather than softened: one widely repeated per-project credit amount for the OpenAI program is not on OpenAI’s own page, so no credit amount is quoted for that row. The ModelRunner row comes from the program page rather than internal knowledge, so it matches what an applicant reads before applying. Terms change without notice — the table is a snapshot, and each program’s own page is the authority.