In 2026, a private startup codebase has typically licensed or sold to AI buyers for about $10,000 to $100,000. Fortune reported in September 2026 that Turing's CEO told The Information it paid an average in the tens of thousands of dollars for each of five to ten failed-startup codebases. A winding-down crypto startup netted about $25,000 for its code (MIT Technology Review, October 2026), and code sold with its tickets, chat and docs can bring more.
On this page
What do AI buyers pay for a codebase?
| Example | Amount | Kind of figure | Source |
|---|---|---|---|
| A qualifying private codebase | Up to $10,000, paid in 7 to 14 days | Post by a writer who works with HUD | DEV Community, June 2026 |
| Five to ten failed-startup codebases | Tens of thousands of dollars each on average | Buyer's CEO | Turing's CEO to The Information, via Fortune, September 2026 |
| A repository of 50,000+ human-written lines | $10,000 to $100,000+ | Buyer's published range | Turing process page, October 2026 |
| A crypto startup's codebase | About $25,000 | Press report of a closed deal | MIT Technology Review, October 2026 |
| One company's 326 repositories | $300,000 | Broker claim, its largest deal to date | Pangea, October 2026 |
| A web app, estimated at published rates | $717 for the code, $5,665 for its commit history | Calculator estimate | aitrainer.work using Pangea rates, October 2026 |
Two patterns stand out. First, history can pay more than code: in the calculator example, the commit history was worth almost eight times the code. Second, the biggest numbers come when code is sold with the rest of a company's record.
The Spirit Airlines archive included 516 repositories, about 30 million lines of code, 372,585 commits and 43,170 pull requests alongside email and chat (The Next Web, August 2026). Google won the August 2026 bankruptcy auction for that archive with a $10 million bid, but the court hearing on the sale was moved to November 6, 2026, so it is not final (TravelPulse, October 2026). Separately, a startup that had just agreed to be acquired received four offers of up to $300,000 for its codebase and employee communications together, and turned them all down (The Information via fnnews, August 2026).
Who buys startup codebases?
Sorted by name. Data Licensors, the company that publishes this page, also licenses company records, including code.
| Buyer | How they buy code |
|---|---|
| aitrainer.work | Points owners to code buyers and runs a value calculator; does not buy code itself |
| Data Licensors | Code as part of a company's records, with secrets removed |
| DataVendor | HUD's marketplace; sellers set prices, and a 20% platform fee applies when earnings are withdrawn (DataVendor, October 2026) |
| Labelbox | One-time purchase of fully defunct startup codebases; its Prism program also seeks IDE session logs and edit traces |
| Mercor | Bid $7.5 million for the Spirit archive, which included code, and offered to buy or license a just-acquired startup's code (The Information via AI Weekly, August 2026) |
| SimpleClosure | Shutdown service with a marketplace for code and workspace data |
| Turing | Buys codebases on their own (50,000+ lines) and as part of a full operating record |
Pangea, a code licensing marketplace, publishes rates of roughly $100 per million tokens of source code and $40 per million tokens of git history (Pangea, October 2026).
What makes a codebase valuable to AI labs?
Labs use real codebases to train and test AI coding agents. Labelbox says what matters is that the system reflects real-world development, with the complexity that comes from building and operating a product with a team (Labelbox, January 2026). In practice buyers look for:
- Private, never public. Anyone can already read public code, so labs pay only for code that was never published (aitrainer.work, October 2026).
- Several contributors, real history. HUD's criteria name multiple contributors and many commits over months or years (DEV Community, June 2026).
- Human-written. Turing does not accept code written by an AI model or pull requests not written by a person, per aitrainer.work (October 2026).
- Unique code, not bulk. Across the codebases SimpleClosure analyzed, about 37% of raw code disappeared once third-party, generated and duplicate code was removed (SimpleClosure, September 2026).
- Less common languages. Roughly half the code SimpleClosure has analyzed is TypeScript or JavaScript, which it says may make less common languages more interesting.
- Exclusivity. Business Insider (October 2026) reported lower demand for a codebase a rival lab already has.
How do I remove secrets from a codebase?
Assume every secret ever committed is still in the history, even if it was deleted later. Before sharing a sample:
- Rotate everything. Revoke and reissue every API key, database password, cloud credential, signing key and webhook secret. A rotated key is harmless even if missed.
- Scan the full history, not just the latest version, with a secret scanner. Check config files, test fixtures, CI files, notebooks and infrastructure files.
- Remove customer data from fixtures, seed files, logs and database dumps.
- Remove personal details from commit messages and comments where needed, such as personal emails or phone numbers.
- Have the buyer scan again. A second pass catches what the first missed. Data Licensors removes passwords, keys, tokens and other secrets, including in code, before anything leaves its hands (Data Licensors, October 2026).
What IP and licensing checks should I do?
The buyer will ask you to promise you own what you are licensing. This is general information, not legal advice; a lawyer can confirm ownership for your company. Check first:
- Assignments. Every employee and contractor should have signed an IP assignment. Code a contractor wrote without one may not be the company's.
- Customer ownership. Custom work built under a services contract may belong to the customer. Read your master service agreements.
- Open-source components. List the third-party and open-source code inside. Licenses such as the GPL carry conditions, and vendored libraries are not yours to license. Most buyers want only your own code anyway.
- Acquisitions and acqui-hires. If an acquirer took the IP, the code may not be yours to sell.
- Who signs. For a closed company, the board, a liquidator, an assignee or a trustee may need to approve. Co-founders and investors may also have a say (aitrainer.work, October 2026).
Our legal guide covers ownership in more detail.
Should I sell outright or license?
A sale is one payment, and the buyer may take broad rights. Labelbox describes its defunct-codebase deals as one-time, with no maintenance or support burden and no public attribution. A non-exclusive license lets you keep ownership and license the same repository to more than one buyer: the HUD post calls its deals "a license, not a sale," and the aitrainer.work calculator notes the same repo can earn again from each lab that licenses it.
If the company is closing anyway, a clean one-time sale is simpler. If you will keep the code, or might restart the product, license it and keep ownership. Either way, read the deletion and use terms in our licensing agreement guide.
How do I start?
- List repositories, languages, years active, contributor count and commit count.
- Rotate credentials and run a secrets scan.
- Confirm ownership and list open-source components.
- Keep the GitHub or GitLab organization paid and active until a buyer has sampled.
- Ask more than one buyer to price the same repositories, with and without the related tickets, chat and docs.
If the whole company is closing, read selling startup data before shutdown too, or compare reported prices in the deal ledger.
