Originally published at aidata.broker.

AI companies do buy private codebases, and they price them by the training tasks a codebase yields, not by its lines. The public prices cluster in a narrow band: a platform that grades repositories starts at a $5,000 baseline per repository, HUD's DataVendor says "About $30,000 per license is the floor for a good codebase", and Turing's March 2026 newsletter priced a modern codebase at $1,000 per qualifying pull request, up to $10K while pricing legacy code (COBOL, .NET, Java) at $10K to $100K+. The middle half of the public per-repository prices runs from $1,478 to $18K per repository, and the realised code sales on record have a median of $1,000 per repository. One closed startup's code base, sold through a wind-down marketplace, netted its founder roughly $25,000, as he told MIT Technology Review.

Watch who buys private code and what it fetches, as a short video, with its transcript.

This guide is for founders, CTOs, CFOs and counsel of software companies, agencies and closed startups that hold private code and are being asked, or are asking themselves, whether it can be sold for AI training. It covers who buys code and how each buyer prices it, why buyers count tasks rather than lines, what raises a codebase's value and what kills a sale outright, how to prepare a codebase, how a codebase valuation works, and a worked example of a typical startup. It sits beside our guide to RL environments, which explains what buyers turn code into.

This guide is research from public sources, not legal, tax or investment advice.

Contents

  1. Who buys private code
  2. Why buyers price by tasks, not lines
  3. What raises a codebase's value
  4. What kills a sale
  5. How to prepare a codebase for sale
  6. How a codebase valuation works
  7. Worked example: a startup codebase
  8. Questions people ask
  9. What this means for you

Who buys private code

Private code reaches AI labs through four kinds of buyer. The first is data vendors that turn code into training tasks and environments for labs: Turing's Project Lazarus, Mercor, micro1 and others. The second is code marketplaces and brokers that license repositories to many buyers: HUD's DataVendor, Pangea, pre.dev and catalogue sites such as Opendatabay. The third is wind-down services that sell a closing startup's assets, code among them: SimpleClosure's AssetHub, Sunset and Labelbox's programme for defunct codebases. The fourth is labs and large technology companies buying directly, which is rarer and mostly visible through benchmarks, bankruptcy dockets and pilot programmes. Our buyer board lists 18 buyers, marketplaces and brokers that take source code.

Each prices differently, and the unit is the first thing to compare. Some price per repository, adjusted for quality; some per pull request or per task the code yields; some per token (a token is a short chunk of text, roughly a word or part of one, the unit models read); some per company, with code as one part of a whole archive. The table puts the published terms side by side; the sections after it explain each.

Buyer Route How it prices Published price Terms worth knowing
Turing (Project Lazarus) Licenses a company's whole operating record to labs; code is welcome but not required Lazarus pays a share of each sale; its March newsletter priced per pull request (modern), per codebase (legacy), per artifact when bundled Newsletter: $1,000 a PR, up to $10K; legacy $10K to $100K+ Paid on each sale; you keep ownership; Turing holds it exclusively for 6 months
DataVendor (HUD) (HUD) Marketplace for non-exclusive licences Per codebase, judged on rights, demand, documentation and difficulty to recreate From about $30,000, its floor for a good codebase Keeps 20% of earnings at withdrawal, and in some countries Stripe's payout fee then too
Pangea.ai Network that licenses code through a partner Per repository by depth, or per token $700 to $1,000 light to $1,500 to $2,000 deep, per licence Paid again on each further licence
pre.dev Buys non-exclusive code licences Per token of source and of git history $6,000 for its example repository Each repository gets its own value and grade
Troveo's cited platform Grades and lists repositories Per repository, adjusted after grading $5,000 baseline Baseline moves up or down with the grade
SimpleClosure (AssetHub) Sells a closing startup's code and workspace data Per company About $25,000 net for one code base 47% of its deals now include code or data
Mercor Buys or licenses code bases with task records Per company, on its listing for software companies licensing private code, more for large codebases or exclusive terms $50K to $500K; it offered to buy Warmly's code base and the records of its tasks Wants repositories, commits, PRs, and issues
Collinear Licenses company data, code included Quotes a range per company No figure published Price rises with history, PRs with tickets, human-written code
Scale AI Bought startups' repositories for a benchmark Not disclosed 18 repositories bought for SWE-Bench Pro Held out as a private test set
Google (Content Offer Pilot) Invites owners of large datasets in specialised segments; code repositories are one example One-to-one offers, fee agreed per partner "such as code repositories", no price published Non-exclusive; owner keeps its IP; the pilot is expanding

Turing's Project Lazarus

Turing published the most detailed price sheet for code. In a March 2026 newsletter it priced modern codebases at "$1,000 per PR, up to $10K per codebase" and legacy codebases at typically $10K to $100K+, depending on complexity, and it offered $50 per artifact for tickets, specs, docs and support threads bundled with a codebase. Its Project Lazarus page, rebuilt by 2 November 2026, states no price for code or anything else: Turing prepares a company's record, licenses it to labs and pays the company every time it sells, and it gives no estimate until it has reviewed a sample. Its chief executive told The Information that Turing had bought five to ten failed-startup codebases, paying an average in the tens of thousands of dollars each.

What matters for a holder is how the licence works. The holder keeps ownership and gives Turing the right to prepare the record and license it, and Turing removes personal data and secrets itself in isolated processing after the holder approves the scope. For the first 6 months after signing, the licence is Turing's alone, and its preparation costs are covered from sales, so ask what share of each sale reaches you and what comes off it first.

Its published screen for code is strict and detailed, and it sits in its technical overview and evaluation script. The evaluator drops pull requests from bot accounts and those not written in English, and counts a pull request only if its linked issue is closed, with a body of 20 to 2,000 words, and it changes at least one test file, with a test that fails before it and passes after it and an earlier test that still passes. Its script also runs a check for signs of AI-generated "vibe coding" in the commit history, without saying what a finding changes. That screen rules out much ordinary product code: changes with no issue and no test behind them, new features with no earlier tests, and possibly code written with an AI assistant.

Whether code alone is wanted is a question to put to Turing. Its FAQ says source code is not required, and welcome when it's there, while its codebase page still has a holder run its evaluation script locally and upload only the CSV it produces for a repository on its own. A holder with nothing but a repository should ask whether Turing will license it alone before preparing for its screen. The gap between the newsletter's cap on a modern codebase and its legacy price is the gap between a modern web codebase and a large or legacy one, and it is the clearest evidence that scarcity, not size, sets a code price: the same buyer's top legacy price is 10x its cap on a modern codebase.

HUD's DataVendor, Pangea and pre.dev

The code marketplaces license a repository to several buyers rather than selling it once. HUD's DataVendor judges "whether your rights are clear, how strong demand is, how well documented the asset is, how hard it would be to recreate, and how directly it fits a buyer's need", says it currently sees strong demand for enterprise software codebases, and asks holders to send a README, setup instructions, ticket history, pull-request context and test information (DataVendor docs). The clean-up is the holder's job there: remove secrets, credentials, customer information and anything you do not have the right to share before connecting or uploading. It keeps 20% of earnings at withdrawal, and in some countries also takes Stripe's payout fee then, a buyer who makes an offer is named to the holder, while one who buys outright or takes a sample stays anonymous, and every sale is final, with no refunds on either side, though a charge the buyer's bank reverses is taken back from the vendor's wallet.

Pangea's code licensing partner quotes the holder's side per repository by depth, from $700 to $1,000 for a light repository to $1,500 to $2,000 for a deep one, per licence, or by the token at $100 per million tokens of source code and $40 per million tokens of git history. On 14 October it said it had paid $600,000 to holders across 380 codebases; on 27 October its referral page said $6M to code owners in the last 60 days, across 11,069 repositories, about $542 a repository. Both are its own counts, which nobody else has confirmed, and the second is the largest record of what code licences have actually paid holders, well under its own per-repository quotes. pre.dev uses the same per-token rates and shows $6,000 per repository for its example, on a non-exclusive licence paid Net 15. The two bases are not the same: DataVendor's figure is before its share, Pangea's is the holder's side with nothing deducted, so compare offers net.

Closed startups

A shut-down startup's code is usually sold together with its Slack, email and documents, by the wind-down service that closes the company. SimpleClosure's test marketplace recovered about $1 million across nearly 100 deals in its first year, about $10K a deal, against early deals reported at $10K to $100K per company. By October 2026 it had handled over 150 deals, and one founder told MIT Technology Review he had netted roughly $25,000 for his code base and expected about $15,000 more for the workspace data, still pending when he spoke. Its terms are worth reading before submitting anything: a deal can be a licence, in which the company keeps the assets, or an outright sale; submitted assets may be acquired by SimpleClosure itself or its affiliates as well as by buyers it introduces; and the company covers SimpleClosure's losses from a breach of its warranties about the assets or from claims that they infringe others' rights, which is one more reason to settle ownership first. Labelbox says it "works with fully defunct startup codebases". At the top of the range, Warmly, after agreeing to be acquired by HubSpot, received four offers of up to $300,000 for its codebase and communication records, and turned them all down. Our guide on selling customer data in bankruptcy and the estates page cover who can sign for a closed company.

Labs buying directly

Labs seldom appear as the buyer of record for a repository, but when they do it shows what they want. Scale AI built the commercial part of its SWE-Bench Pro benchmark from 18 proprietary repositories of early-stage startups, "where we contacted the company and purchased their engineering repos".

Google runs an open invitation, and it is for content in general rather than code alone. Its Content Offer Pilot invites any owner of large, high-quality datasets in specialised content segments, such as code repositories, to express interest. On 18 October its page said the pilot is now expanding to accommodate increased demand, after more than a year of testing with thousands of partners. Its terms are among the most open on our buyer board: non-exclusive, the owner keeps 100% of its IP and may sell the same content elsewhere, and the fee is agreed with each partner in an offer process that starts on invitation. Two limits matter for code. The page publishes no price, and it does not say the content trains models: partners are paid for content that will be used to improve Google's products and services, including quality evaluations, product features and display across its products. A holder should ask for the use to be named in the licence. In June 2026 Google also emailed a select group of Google Play app developers inviting them into the pilot; our guide to Google's Content Offer Pilot reads that email, the pilot's terms and how its non-exclusive licence meets other buyers' clauses.

In the Spirit Airlines bankruptcy, Google won the auction for the airline's business data and software code for $10 million, a sale still awaiting court approval as of this writing.

Prices on record

What Price Unit Date Source
A private repository graded by a platform $5,000 baseline, adjusted after grading per repository Sep 2026 Troveo
A good codebase on HUD's DataVendor From about $30,000 per licence (its floor) Nov 2026 DataVendor docs
A modern codebase, priced by pull request (Turing) $1,000 per PR, up to $10K per codebase Mar 2026 Turing
A legacy codebase: COBOL, .NET, Java (Turing) $10K to $100K+ per codebase Mar 2026 Turing
A company's record, code included, licensed by Project Lazarus A share of every sale, no fixed price stated per sale Nov 2026 Turing
A private repository through Pangea, light to deep $700 to $2,000 per repository, per licence Oct 2026 Pangea
pre.dev's example repository $6,000 per repository, non-exclusive Oct 2026 pre.dev
A failed startup's codebase (Turing bought several) Tens of thousands of dollars per codebase Sep 2026 Fortune
A closed crypto startup's code base (SimpleClosure) About $25,000, net to the founder per code base Oct 2026 MIT Technology Review
A startup's code and communication records (Warmly) Up to $300K (four offers, refused) per company Aug 2026 The Information, via FNNews
One enterprise repository listed for sale on Opendatabay $65,800 asking price per repository Sep 2026 Opendatabay
A task built from a contributor's own private codebase $300 per codebase plus $75 per approved task per codebase and task Oct 2026 AfterQuery on Handshake

Sources: Pangea, light, Pangea, deep, graded platform, pre.dev example, Turing, modern (one PR to its cap), HUD DataVendor (its stated floor). Turing's legacy prices of $10K to $100K+ are left off the chart because they would flatten every other bar.

Two warnings about these numbers. Most are ceilings or offers, not averages of closed deals: SimpleClosure's realised average of about $10K a deal sat at the bottom of the range it quoted, and other intermediaries show the same pattern (headline ceilings are not what deals pay). And they price different things: a light repository on a per-licence basis is not the same asset as a fifteen-year legacy codebase sold once. The source code page keeps the current anchors.

The companies that build coding data for AI labs, in order of the money on record, are on our ranking of coding data companies.

Why buyers price by tasks, not lines

Public code is not scarce. The Software Heritage archive holds over 27 billion unique source files, the open Stack v2 dataset over 3 billion files in 600+ languages, and GitHub hosts 630M repositories. Models are already trained on code like that. A line of private code that looks like a line of public code teaches a model almost nothing new, which is why one buyer's guide puts it bluntly: "A snapshot of source code with no history is worth little to a lab, because there's already a lot of code in the world."

What public code lacks is the record of how work was done: the request, the change that answered it, the review that argued over it, and the test that proved it. Labs train coding agents on exactly that unit, a task: the repository frozen before a change, the issue that asked for it as the prompt, the merged change as the reference answer, and the tests as the verifier. Our guide to RL environments explains how such tasks are used. A codebase is worth what it yields in those tasks, which is why Turing prices modern code per pull request and why buyers of company records say the same thing about links: Sunset's chief executive told Forbes that a Jira ticket tied to a specific code commit is worth more than a standalone document (Forbes, via Forbes Japan).

Most real engineering work happens out of public view, and that is where the opportunity sits: GitHub reported that 81.5% of contributions in 2025 happened in private repositories. But the same logic cuts the other way on volume. SimpleClosure found that roughly 37% of raw code disappeared once third-party code, generated code and internal duplicates were removed, and its chief executive says a codebase a quarter of the size can be worth double. Net, human-written, linked code is what gets counted.

There is also a ceiling set by machines. Together AI generated 258,000 test-verified coding-agent trajectories for $130k in compute, about 50 cents each. What a model can generate for itself is worth about what it costs to generate. What it cannot generate is real engineers' changes to real private systems, reviewed and tested over years.

What a buyer counts Why it matters
Merged pull requests that resolve one linked, closed issue Each is a task with a prompt (the issue) and a reference answer (the change)
Tests that fail before a change and pass after it, run in CI The tests are the verifier; code without tests yields few tasks
Review comments on each change They record how engineers judged quality, which buyers train reviewers on
Years of continuous history More changes, more kinds of change, and the evolution of a real system
Code that builds in a container A buyer must run every task as delivered; Turing counts a change only with a test that fails before it and passes after it
Code written by people AI-written code teaches a model to copy models; Turing's script checks the commit history for AI-generated code
Code that was never public A model cannot have seen it, so it is uncontaminated

What raises a codebase's value

The buyers that publish criteria name the same few things, and our valuation weights each by how many code buyers name it in their own words. Tests that run in CI come first, because the tests are the verifier for every task built from the code; Turing counts a pull request only if it changes at least one test file, and HUD asks for test information with every codebase. Reviewed changes tied to their tickets come next: Turing asks for "Source code, reviews, tickets, and the iteration history behind what shipped", Mercor for "GitHub Repositories, commits, PRs, and issues", and Collinear says its price rises with pull requests with their tickets, several years of history, human-written code and the same work visible across several tools.

Scarce languages and legacy systems carry the largest premium on record. Turing's legacy price for COBOL, .NET and Java is 10x its modern cap, and SimpleClosure reports that less common languages such as Ruby or C++ can fetch top dollar, better still with Jira, Linear or Figma context. Legacy code is scarce in training data and abundant in the world: one vendor survey estimated 775 to 850 billion lines of COBOL in daily use, which is why agents that can maintain it are worth training.

Years of history and an unusual domain raise the price because they hold more, and more varied, real work. Troveo says "Value scales with your systems, team size, and years of history.", and AfterQuery prices "by depth, breadth and uniqueness", including how hard the record would be to recreate. A fintech ledger, a logistics routing engine, industrial control software or a scientific code base teaches more than another web app.

The company's other records alongside the code raise it further. Turing has offered $50 per artifact for tickets, specs, docs and support threads bundled with a codebase, and whole operating records with code inside sell per company rather than per repository: micro1's partnership tiers run from $100K-500K to $3M+, and Turing licenses a whole operating record to labs and pays the company on every sale. A company that sells its code alone leaves that on the table; our software companies page lists what buyers ask software companies for.

Raises value Lowers value
Tests running in CI, a Dockerfile, a passing main branch No tests, or code that does not build
Pull requests with review comments, each closing a ticket Direct pushes to main with one-line commit messages
Commits that name their ticket keys Tracker and code kept apart, with no keys to join them
Scarce languages, legacy systems, unusual domains Common web and CRUD apps many holders offer
Years of continuous history from a real team A short history, or a single squashed commit
Mostly written by people, much of it before AI assistants A large share of commits co-authored by AI assistants
Never public Forks, open-sourced modules, published packages
Tickets, docs and Slack sold with the code Code alone

What kills a sale

Some problems lower a price. Others stop a sale until they are fixed, because no responsible buyer will take the code and no holder should warrant it. Buyers' contracts make the holder warrant that it owns what it licenses, and buyers have been burned: Surge AI's head of environments described being offered a large company's codebase for millions of dollars by a group that had "illicitly obtained a copy of the codebase". The checks below are the ones that most often stop a code sale. They are why a rights file, the written record of who owns what, comes before any conversation about price.

Open-source licence obligations

Most codebases contain open-source code, and its licence travels with it. Permissive licences (MIT, Apache, BSD) mostly ask for attribution and a copy of the licence. Copyleft licences such as the GNU GPL and the Source:https://dev.to/aidata_broker/how-to-sell-your-codebase-or-source-code-to-ai-companies-38nc