AI companies license private source code together with its full history of commits and pull requests, mostly through company data programs, RL-environment builders and code marketplaces rather than by buying it outright. The page covers code as one data type in AI training data licensing: who handles it, which rates they state, which codebases qualify and which rights the owner clears first.
Pangea.ai states indicative rates of USD 100 per million tokens of source code and USD 40 per million tokens of git history (checked 2026-11-09), and a bankruptcy schedule counts 516 repositories in a data package auctioned but not approved for sale as of 2026-11-09. DataPacts takes no cut of any license, and not licensing is a valid outcome.
Every figure here is a claim by the party named, with its check date, not a paid amount. The method is explained under how we source, date and check every figure.
What Code AI Companies License and Why the History Matters
AI companies want code that shows how real teams build and change software over years, so the git history is licensed together with the source and is priced as its own unit. Pangea states separate rates for source and for history, and pre.dev states that it quotes per-token pricing for source and history separately on a call (checked 2026-11-09).
Acquirers describe the same four traits:
- Private repositories that were never public
- Long commit and pull-request history
- Several human authors
- Tests that run
The one court record that counts code in an AI data sale is the Spirit Airlines asset schedule (Bankr. S.D.N.Y. 25-11897, Dkt. 1463 Exh. A-2): 516 repositories, 43,170 pull requests and 372,585 commits, all marked Included. The bids cover the whole de-identified data package, not the code alone: Google LLC's successful bid is USD 10,000,000 and the Mercor.io Corporation alternate bid is USD 7,500,000, from the auction of 2026-08-14. The sale is not approved as of 2026-11-09, with a hearing set for 2026-11-17 (Dkt. 1779).
Code is one family among the data types AI companies buy, alongside workplace records, media and regulated records.
Who Licenses Codebases: Programs, RL-Environment Builders and Intermediaries
Fourteen companies state on their own pages that they license, resell or broker private code for AI training, and only some of them take the license themselves. The table below lists them alphabetically as of 2026-11-10, with terms checked on 2026-11-09 and each party's own page as the source.
- Role is the type used on this site: program, data vendor, RL-environment builder or intermediary.
- Intermediaries (marketplaces, brokers, shutdown services) handle, list or match code for others and do not buy it.
- Figures are each party's claim and are never added or compared across rows.
| Acquirer or intermediary | Role | What it acquires or handles | Figure or pricing basis it states | Source |
|---|---|---|---|---|
| Abaka AI (Project Reborn) | data vendor | private codebases with full commit history | up to USD 1M per repo, up to USD 5M per company | abaka.ai |
| AfterQuery | human-data platform | codebases among company records | no payout figures published | afterquery.com |
| aitrainer.work | broker (intermediary) | private code repos, matched to buyers | buyer sets price | aitrainer.work |
| Collinear AI | RL-environment builder | private repos with PRs and tickets | no amounts; paid per contract after delivery | collinear.ai |
| DataVendor (HUD) | marketplace (intermediary) | private codebases listed by vendors | floor about USD 30,000 per codebase license; 20 % fee at withdrawal | docs.datavendor.ai |
| Mercor (Software Company listing) | human-data platform | private repos with 3+ years of history | USD 50,000 to 500,000 to the company | api.data.mercor.com |
| Origin Lab | data vendor | game source code and commit history | USD 100 to 100,000 per game | originlab.ai |
| Pangea.ai | marketplace (intermediary) | private source code and git history | indicative USD 100 and USD 40 per million tokens | pangea.ai |
| Patronus AI | RL-environment builder | production codebases with history | no price; written offer after a sample | data.patronus.ai |
| pre.dev | RL-environment builder | private production codebases with full git history | per-token pricing quoted on a call; seller payout not published | pre.dev |
| Replay | data vendor | code and commits among company records | upfront payment plus perpetual revenue share; bands by headcount, not code-specific | replay.ai |
| SimpleClosure (AssetHub) | broker, shutdown service (intermediary) | shutting-down startups' code and workspace data | percentage only on success; "tens to hundreds of thousands of dollars for unique assets" | simpleclosure.com |
| Sureform | RL-environment builder | codebases, PRs and issues | no payout range on its site | sureformhq.com |
| Turing (Project Lazarus) | human-data platform | source code (GitHub, GitLab) and reviews | revenue share, nothing upfront (Lazarus page); per-PR prices in newsletter 57 | lazarus.turing.com; turing.com |
None of the 14 pages states a frontier lab's purchase price for a codebase. The SpaceX and Cursor agreement of 2026-04-19 is a compute and option agreement covering developer interaction data, not a data price (SEC filing, SpaceX).
Per-Token, Per-PR and Per-Codebase Rates on Record
Code is priced in three units on the record as of November 9, 2026: per million tokens, per pull request and per codebase or repository, and every figure is the stating party's claim, not a paid amount.
The table lists each stated code rate as of 2026-11-10, one row per claim, none averaged or combined.
| Unit | Figure | Stating party (role) | Statement date | Checked | Source |
|---|---|---|---|---|---|
| per million tokens, source code | USD 100, indicative; can change on 30 days' notice, never retroactively | Pangea (marketplace) | undated page | 2026-11-09, re-read 2026-11-10 | pangea.ai/code-licensing/companies |
| per million tokens, git history | USD 40, indicative | Pangea (marketplace) | undated page | 2026-11-09, re-read 2026-11-10 | pangea.ai/code-licensing/companies |
| per token, source vs history | quoted on a call, not published | pre.dev (RL-environment builder) | 2026-09-27 page | 2026-11-09 | pre.dev/labs |
| per pull request, modern codebase | USD 1,000, up to USD 10,000 per codebase | Turing (newsletter 57) | 2026-03-25 | 2026-11-09 | turing.com/blog/agi-advance-newsletter-57 |
| per legacy codebase | USD 10,000 to 100,000+ | Turing (newsletter 57) | 2026-03-25 | 2026-11-09 | turing.com/blog/agi-advance-newsletter-57 |
| revenue share, Project Lazarus | nothing upfront; split not published | Turing (Lazarus page) | current page | 2026-11-09 | lazarus.turing.com |
| per codebase deal, Software Company listing | USD 50,000 to 500,000 to the company | Mercor (human-data platform) | current listing | 2026-11-09 | api.data.mercor.com |
| per repository / per company | up to USD 1,000,000 / up to USD 5,000,000 | Abaka AI (data vendor) | current page | 2026-11-09 | abaka.ai/pages/project-reborn |
| per codebase license | floor about USD 30,000; "some single-codebase deals exceeded USD 10 million" | DataVendor (marketplace) | current docs | 2026-11-09 | docs.datavendor.ai |
| largest deal to date | USD 300,000 for 326 repositories from one company | Pangea (marketplace) | undated | 2026-11-09 | pangea.ai/code-licensing/referral |
| net proceeds, one startup code base | roughly USD 25,000 | a founder, to MIT Technology Review | 2026-10-13 | 2026-11-09 | technologyreview.com |
Turing states two models that contradict each other. Its newsletter 57 of 2026-03-25 gives per-PR and per-codebase prices, while its Project Lazarus page states a revenue share with nothing upfront (checked 2026-11-09). Both stand as claims, and neither is picked. No acquirer states how it counts tokens, so a per-token rate cannot be turned into a price per repository without an assumption.
A hypothetical codebase with 20 million source tokens and 60 million git-history tokens shows the arithmetic: 20 x USD 100 plus 60 x USD 40 gives USD 4,400 at Pangea's indicative rates of 2026-11-09. That is not a quote and not a typical price. Rates for other data types are compared under AI training data prices.
What Qualifies: Acquirers' Eligibility Checks for a Codebase
Acquirers that publish eligibility rules exclude public and forked code, and the stricter ones add human-written code, several years of commit history and several engineers. The table compares the published rules as of 2026-11-10, checked 2026-11-09, alphabetically.
| Acquirer or intermediary | Code it accepts | Code it rejects or strips | Source |
|---|---|---|---|
| Abaka AI | 500+ PRs, 3+ contributors, handwritten code from 2023 or earlier, full commit history | not stated beyond the minimums | abaka.ai |
| aitrainer.work | assets the seller owns or holds written permission for from each contributor | not stated | aitrainer.work |
| Collinear AI | several years of human-written history, teams of about 50+, English data; US, CA, UK, EU, AU, NZ | not stated | collinear.ai |
| DataVendor (HUD) | private codebases from organizations | public repositories ("cannot be estimated") | docs.datavendor.ai |
| Mercor | private repos never public, mirrored or open-sourced; 3+ years of commits from 10+ engineers; 50,000+ lines owned outright; human-written code with tests that run | forks, student, hackathon or side projects, code mostly written by AI tools, client work an agency does not own, single-author repos | api.data.mercor.com |
| Pangea.ai | owned private code only | client-owned, public or forked code | pangea.ai |
| Patronus AI | codebases with owner authority | contractor IP assignments asked | data.patronus.ai |
| pre.dev | all branches and tags | public, forked, archived or published-package code; secrets and personal data removed; authors pseudonymized; vendored dependencies, binaries and build output stripped | pre.dev |
What fails is consistent across the stricter pages: Mercor lists single-author repositories, client work an agency does not own, and student or side projects as not a fit (checked 2026-11-09).
Clearing a Codebase Before Licensing: Authorship, Host Terms and Components
Acquirers make the code owner warrant that it owns what it licenses: Pangea.ai's ownership warranty is uncapped, aitrainer.work asks for ownership or written permission from each contributor, and Patronus AI asks for contractor IP assignments.
Contributor Authorship and Assignments
Code written by employees in the course of their work usually belongs to the employer, while code from contractors and freelancers needs a written assignment or grant. The rule rests on 17 U.S.C. 201(b) and 101 in the US, CDPA 1988 s.11(2) in the UK and UrhG § 69b in Germany, each subject to a contrary agreement. The detail is under who owns the code you want to sell.
Code Host Terms: GitHub's Ownership and License Clauses
Hosting code on GitHub does not transfer it: GitHub's Terms of Service, effective 2026-04-27, state in section D.3 "You own Your Content".
Section D.4 gives GitHub a license that covers training its own AI features and models and is stated not to be a sale. A GitHub organization migration archive is automatically deleted after seven days, as GitHub's migrations documentation states.
Third-Party and Open-Source Components in a Licensed Codebase
Acquirers strip or reject code the owner did not write: Mercor rules out forks and copies of open-source projects, pre.dev strips vendored dependencies, binaries and build output, and SimpleClosure states that roughly 37 % of raw code offered through AssetHub is removed in cleaning (blog of 2026-09-03). Whether a license condition on a component affects a code license is covered under open-source license contamination.
How a Code License Runs: Access, Cleaning, Terms and Payment
Most code licenses start with read-only access to the repositories, continue with automated grading and scrubbing, and end with a non-exclusive license and payment after the acquirer or its buyer has paid or accepted.
Access comes first. Pangea asks the owner to connect GitHub, GitLab or Bitbucket read-only, and Origin Lab states read-only code access for games. Abaka AI runs a local metadata-only scan and states a written offer within 14 days (checked 2026-11-09).
Cleaning follows. Pangea states grading and scrubbing before batches go to buyers over 2 to 16 weeks, and pre.dev states that secrets and personal data are removed and authors pseudonymized. Secrets and credentials left in commit history are the owner's risk until that step is done.
Terms are mostly non-exclusive: Pangea states that the owner keeps ownership and that completed licenses survive withdrawal, DataVendor states that a sale licenses rather than transfers, and Collinear AI states a one-time snapshot. Turing states a 90-day window in which it alone sells the record. The trade-offs are covered under exclusive vs non-exclusive data licenses. Pangea requires 30 days' notice to withdraw once code has been presented, and DataVendor states that the buyer's entitlement survives unpublishing.
Payment follows the buyer. Pangea states payment within 15 days of the buyer's payment, Mercor states a wire typically 2 to 4 weeks after signing with its costs covered, and Abaka AI states payment in two parts, the second after quality checks. DataVendor keeps 20 % of earnings at withdrawal (all checked 2026-11-09). Pangea also states that licensing is not guaranteed.
Licensing a Codebase: Benefits, Risks, Downsides and Alternatives, Including Not Licensing
Licensing a codebase turns an asset that already exists into income without giving up ownership, but the code cannot be recalled from a model once it is trained on it.
Benefits
- Ownership stays with the company in every non-exclusive license on record (Pangea, Turing's Project Lazarus, DataVendor, Collinear AI, checked 2026-11-09).
- Idle or legacy code can earn: Turing states USD 10,000 to 100,000+ per legacy codebase in its newsletter of 2026-03-25, and Abaka AI asks for handwritten code from 2023 or earlier.
- Some programs cover the cost of scrubbing: Mercor covers extraction, anonymization and transfer, and Turing covers cleanup and packaging.
Risks
- The ownership warranty can be uncapped, as Pangea states, so a contributor's claim falls on the licensor.
- Secrets and credentials in commit history reach the acquirer unless removed, which pre.dev's stripping step is designed to handle.
- Releasing controlled technology or source code to a foreign person in the US is a deemed export under 15 CFR 734.13, a point covered under export controls on technical data.
Downsides
- Sale is not assured: aitrainer.work states "Most repos don't sell at all" (checked 2026-11-09).
- Cleaning shrinks the offer: SimpleClosure states that roughly 37 % of raw code is removed in cleaning.
- A non-US owner faces 30 % US withholding on royalties under 26 U.S.C. 1442(a) and 881(a), reduced only by treaty, with software following IRS income code 12 (Table 1, rev. May 2023).
Alternatives
- License some repositories, or source without history, as Pangea's separate rates allow.
- Build RL tasks from commits, as pre.dev and Collinear AI describe.
- Use a shutdown sale through a service such as SimpleClosure, if the company is closing.
- Do not license at all and keep the code private.
The question of whether any license makes sense is answered under whether to sell or license your data, which applies the same benefits and risks to every data type.
Where Code Licensing Meets Systems, Startups, Agent Data and the Seller Guide
Code is one data type among the records a company can license; the export route, the seller's situation and the seller guide decide the rest.
- How to sell or license company data: the full seller guide.
- Rights audit for every data type.
- Export rights by source system for code hosts and other systems.
- PII removal and de-identification before licensing.
- Computer-use and agent trajectory data from AI coding products.
- How to sell an RL environment built from real commits.
- Selling data from outside the US for non-US code owners.
- AI training data marketplaces compared on fees and eligibility.
- Referral programs of AI data buyers for those who introduce code owners.
- GitHub repositories: export rights and licenses.
- Software companies licensing code and product data.
- Venture-backed startups licensing data: investors, code IP and exit effects.
- Cursor and SpaceX AI agreement: the developer interaction data agreement.
- Doe v. GitHub (Copilot): dismissal of the DMCA 1202(b) claim affirmed on 2026-09-16 (9th Cir. No. 24-7700, as of 2026-11-09).
This page is information, not legal advice.
Frequently Asked Questions
Can you sell code to AI companies?
Yes, private code the company owns is licensed through programs and intermediaries, usually non-exclusively. Selling scripts or themes on code marketplaces is a different market. Programs that take code are compared under company data partnership programs.
How much do AI companies pay for a codebase?
There is no rate card; stated figures run from per-token rates to per-codebase ranges, each a claim. Pangea states USD 100 per million source tokens and Mercor states USD 50,000 to 500,000 (checked 2026-11-09). Methods are explained under what company data is worth to AI companies.
Is there a code value calculator?
No acquirer publishes one on record. Pangea's per-token rates allow arithmetic only once tokens are counted, and no acquirer states a counting method. Published calculators for other data are run under buyer estimators side by side.
Can legacy code be licensed for AI training?
Yes. Turing's newsletter of 2026-03-25 states USD 10,000 to 100,000+ per legacy codebase, and Abaka AI asks for handwritten code from 2023 or earlier (checked 2026-11-09).
Can a shut-down startup license its codebase?
Yes, through whoever still holds authority over the assets. A founder told MIT Technology Review on 2026-10-13 that a code-base sale netted roughly USD 25,000. The process is under selling data from a shut-down or defunct company.
Can developers get paid to write or review code for AI training?
That is contributor work, paid per task or hour, not a code license. The work is individual and not tied to a company-owned codebase. Pay and assessments are compared under AI training jobs and platforms.