Rightmark Price Index, Q3 2026 edition
Private data now has a public price.
What AI developers pay for data, by type and unit, from public deals and buyer programs. Every price is sourced and dated.
Rightmark Price Index
Q3 2026 edition
Editions
- Q3 2026CurrentPublished 30 Sep 2026, updated 2 Oct 2026
- Next editionDue January 2027
This is the first edition. Each new edition is dated, with its notes.
- Coverage
- 23 prices from 21 sources
- Published
- Updated
- Next edition
- Private repository with git historyper repo, per sale$150 to $1,500Medium confidenceRightmark summary of public code licensing programs, Q3 2026
- $200 to $2,000Medium confidenceEpoch AI, Jan 2026
- Shut-down startup archiveper company, code and workspace$10,000 to $100,000Medium confidenceGizmodo, Apr 2026
- $185,000 to $1M+Your range $230,000 to $440,000Medium confidenceRightmark summary of public company data programs, 2026
- Large enterprise archiveper company, winning bid, pending court approval$10MMedium confidenceTIME, Aug 2026
- Unused creator footageper minute$1 to $4Medium confidenceThe Decoder, Jan 2025
- Low-resource language speechper hour$60 to $150Medium confidenceSpeechData.ai, Jun 2026
- Nonfiction backlistper title, for 3 years$5,000High confidenceeWeek, Dec 2024
8 of 23 prices, chosen to span the market. Each row keeps its own unit, plotted on one log scale from $1 to $10M to show the spread. Your appraisal is the orange range marked You, on the Operating company, 30+ staff, English records row.Your range $230,000 to $440,000. Indicative estimate, not an offer. Final offers depend on the systems you connect, data size and quality, access terms and due diligence.
Show these prices as a table
| Data type | Price | Unit | Confidence | Source |
|---|---|---|---|---|
| Private repository with git history | $150 to $1,500 | per repo, per sale | Medium confidence | Rightmark summary of public code licensing programs, Q3 2026 |
| Training task built from real code | $200 to $2,000 | per task | Medium confidence | Epoch AI, Jan 2026 |
| Shut-down startup archive | $10,000 to $100,000 | per company, code and workspace | Medium confidence | Gizmodo, Apr 2026 |
| Operating company, 30+ staff, English records | $185,000 to $1M+Your range $230,000 to $440,000 | per company | Medium confidence | Rightmark summary of public company data programs, 2026 |
| Large enterprise archive | $10M | per company, winning bid, pending court approval | Medium confidence | TIME, Aug 2026 |
| Unused creator footage | $1 to $4 | per minute | Medium confidence | The Decoder, Jan 2025 |
| Low-resource language speech | $60 to $150 | per hour | Medium confidence | SpeechData.ai, Jun 2026 |
| Nonfiction backlist | $5,000 | per title, for 3 years | High confidence | eWeek, Dec 2024 |
The Rightmark Price Index records what AI developers pay for data, by type and unit, from public deals and buyer programs. In Q3 2026: $150 to $1,500 per code repository per sale, $1 to $4 per minute of unused video, $50 to $150 per hour of speech, $5,000 per book title, and a $10M winning bid for one company archive.
- Code repositories: $150 to $1,500 per repo per sale. Rightmark,
- Unused video footage: $1 to $4 per minute. The Decoder, (opens in a new tab)
- Nonfiction backlist: $5,000 per title for 3 years. eWeek, (opens in a new tab)
- A large company archive drew a $10M winning bid in a bankruptcy auction, pending court approval. TIME, (opens in a new tab)
Updated
Rightmark Price Index, Q3 2026: what AI developers pay for data, by data type and unit, with confidence and source
- High confidence
- Medium
- Low, inferred
- Midpoint of a sourced range
- Your appraisal
23 of 23 prices. Log scale, each price per its own unit. Pick one unit to compare rows directly. Your appraisal is the orange bar on Operating company, 30+ staff, English records.
| Range on a log scale, US dollars | Confidence and source | Details | ||
|---|---|---|---|---|
| Code and repositories3 prices, per repo, per taskExplore code | ||||
| Private repo, standing-rate licenseTypical, per sale. Large repos up to about $10,000 per sale (indicative); history often carries more than half the price | $150 to $1,500per repo | Midpoint of a sourced range: $474 per repo. | Medium confidenceRightmark summary of public code licensing programs, Q3 2026 | |
What this price coversA non-exclusive license to one private repository with its full git history, priced per token of original code and, at a lower rate, per token of git history, per sale. Large repos with long histories reach about $10,000 per sale (indicative), and each further license pays again. Midpoint of a sourced range: $474 per repo. What moves it
SourceRightmark summary of public code licensing programs, Q3 2026 No public link: Rounded by Rightmark from pre-sale estimates shown in public code licensing programs in September 2026; no single public page states this range. Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Premium or task-grade repositoryOur estimate for repos that build, pass their tests and carry real bug-fix history; no published per-repo price | $10,000 to $100,000+per repo | Midpoint of a sourced range: $31,600 per repo. | Low confidenceRightmark estimate for task-grade code repositories, Q3 2026 | |
What this price coversRepositories that build, pass their tests and carry real bug-fix history, which buyers can turn into verifiable training tasks. No buyer publishes a per-repo price for these, so this range is our estimate. Midpoint of a sourced range: $31,600 per repo. What moves it
SourceRightmark estimate for task-grade code repositories, Q3 2026 No public link: Our own estimate for repos that build, pass their tests and carry real bug-fix history; no published per-repo price exists. Low confidenceInferred, implied or estimated, so read it as rough. | ||||
| Training task built from real codeReported range per task; difficulty and fair, repeatable grading set the price | $200 to $2,000per task | Midpoint of a sourced range: $632 per task. | Medium confidenceEpoch AI (opens in a new tab) Jan 2026 | |
What this price coversOne training task built from real software work: a problem with a checkable answer, used to train or test a model. Midpoint of a sourced range: $632 per task. What moves it
SourceEpoch AI, Jan 2026 (opens in a new tab)An FAQ on Reinforcement Learning Environments Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Operational data3 prices, per companyExplore company data | ||||
| Shut-down startup, code and workspaceReported sales of closed startups' data | $10,000 to $100,000per company | Midpoint of a sourced range: $31,600 per company. | Medium confidenceGizmodo (opens in a new tab) Apr 2026 | |
What this price coversCode plus workspace data, such as chat, email and docs, from a startup that has stopped trading, sold as one archive. Midpoint of a sourced range: $31,600 per company. What moves it
SourceGizmodo, Apr 2026 (opens in a new tab)Failed Companies Are Selling Old Slack Chats and Email Archives to Train AI Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Operating company, 30+ staff, English records30 to 99 staff typically $185,000 to $720,000, depending on years and region; 100 to 499 staff reach the low millions. Lump sums are usually exclusive for a term | $185,000 to $1M+per companyYour appraisal$230,000 to $440,000 | Midpoint of a sourced range: $430,000 per company. | Medium confidenceRightmark summary of public company data programs, 2026 | |
What this price coversEmail, chat, documents, tickets, CRM records and code from a company with 30 or more staff and mostly English records, licensed as one archive. Lump sums are usually exclusive for a term; a non-exclusive revenue share is the alternative. Midpoint of a sourced range: $430,000 per company. What moves it
SourceRightmark summary of public company data programs, 2026 No public link: Rounded by Rightmark from published price calculators and stated ranges for exclusive lump-sum programs; no single public page states these bands. Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Large enterprise archiveWinning bankruptcy-auction bid for Spirit Airlines' archive (about 100M emails, 500M Teams messages, 30M lines of code); the sale still needs court approval and a rival bidder later offered $12.5M | $10Mper company | Medium confidenceTIME (opens in a new tab) Aug 2026 | ||
What this price coversA whole company's archive, with its email, Teams chat and code, offered in a bankruptcy auction. Google's $10M bid won the auction, but the sale still needs court approval, so read it as a winning bid rather than a completed price. What moves it
SourceTIME, Aug 2026 (opens in a new tab)Why Google Bid $10 Million for a Failed Airline’s Data Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Voice and audio3 prices, per hourExplore audio | ||||
| Bulk call-center audioMarketplace asking price, 1M-hour multilingual set | $6.30 to $7per hour | Midpoint of a sourced range: $6.64 per hour. | Medium confidenceDatarade via Internet Archive (opens in a new tab) Aug 2025 | |
What this price coversThe asking price for a 1M-hour multilingual set of call-center audio on a data marketplace. Transcribed, labeled and consented sets sell higher. Midpoint of a sourced range: $6.64 per hour. What moves it
SourceDatarade via Internet Archive, Aug 2025 (opens in a new tab)Global Call Center & Conversational Audio Dataset, Multilingual, Validated, with Demographics + Custom Collection Available Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Ready-made speech datasetsCustom collection runs $150 to $400+ | $50 to $100per hour | Midpoint of a sourced range: $71 per hour. | Medium confidenceSpeechData.ai (opens in a new tab) Jun 2026 | |
What this price coversOff-the-shelf, transcribed speech in major languages, priced per hour of audio. Custom collection runs $150 to $400 or more per hour. Midpoint of a sourced range: $71 per hour. What moves it
SourceSpeechData.ai, Jun 2026 (opens in a new tab)How to Buy Speech Data: The Complete Buyer's Guide Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Low-resource language speechPrice rises with language rarity | $60 to $150per hour | Midpoint of a sourced range: $95 per hour. | Medium confidenceSpeechData.ai (opens in a new tab) Jun 2026 | |
What this price coversReady-made, transcribed speech in low-resource languages, which sits at the top of the speech range. Midpoint of a sourced range: $95 per hour. What moves it
SourceSpeechData.ai, Jun 2026 (opens in a new tab)How to Buy Speech Data: The Complete Buyer's Guide Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Video and imagery4 prices, per minute, per hour, per imageExplore video and images | ||||
| Unused creator footage4K, drone and 3D at the top | $1 to $4per minute | Midpoint of a sourced range: $2 per minute. | Medium confidenceThe Decoder (opens in a new tab) Jan 2025 | |
What this price coversFootage that creators shot but never published, licensed to video model labs by the minute. Standard 1080p sits around $1 to $2 a minute; 4K, drone and 3D reach $4. Midpoint of a sourced range: $2 per minute. What moves it
SourceThe Decoder, Jan 2025 (opens in a new tab)OpenAI and Google are buying YouTubers' unpublished videos for up to $4 per minute Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Stock footage bought by a major platformReported purchase program | $2.60 avg, up to $7.25per minute | Medium confidencePetaPixel (opens in a new tab) Apr 2024 | ||
What this price coversA major platform's reported program to buy videos for AI training, which paid about $2.60 a minute on average and up to $7.25. What moves it
SourcePetaPixel, Apr 2024 (opens in a new tab)Adobe Will Buy Your Videos for Up to $7.25 Per Minute to Train AI: Report Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| First-person task videoLabels, hands in frame and sensors raise it | $15 to $60per hour | Midpoint of a sourced range: $30 per hour. | Medium confidenceStellaris VC (opens in a new tab) Apr 2026 | |
What this price coversVideo of real tasks filmed from the worker's point of view, used to train robotics foundation models. Unlabeled footage sits at the low end. Midpoint of a sourced range: $30 per hour. What moves it
SourceStellaris VC, Apr 2026 (opens in a new tab)Physical AI Has a Massive Data Problem Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Stock imagesCaptions and releases set the price | $0.02 to $2per image | Midpoint of a sourced range: $0.20 per image. | Medium confidenceReuters via Rappler (opens in a new tab) Apr 2024 | |
What this price coversReported license prices for images in AI training deals, from 2 cents to about $2 per image. Midpoint of a sourced range: $0.20 per image. What moves it
SourceReuters via Rappler, Apr 2024 (opens in a new tab)Inside Big Tech's underground race to buy AI training data Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Health and life sciences2 prices, per record, per slideExplore health | ||||
| De-identified EHR, imaging, pathologyNo buyer publishes per-record prices | Revenue shareper record | No public per-unit price | Low confidenceSTAT (opens in a new tab) Jan 2025 | |
What this price coversDe-identified records, imaging and pathology from hospitals and clinics, licensed to health AI developers. No buyer publishes a per-record price, and deals are priced as a revenue share. What moves it
SourceSTAT, Jan 2025 (opens in a new tab)The companies paying hospitals to hand over patient data to train AI Low confidenceInferred, implied or estimated, so read it as rough. | ||||
| Pathology slides (acquisition-implied)Implied by a 2025 acquisition that also bought products and models | About $11.60per slideImplied by the reported deal in its source, not a price for this data alone | Low confidenceBusinessWire (opens in a new tab) Aug 2025 | ||
What this price coversA per-slide figure implied by a 2025 acquisition that cited about 7M annotated slides. The deal also bought products and models, so it is not a price for slides alone. What moves it
SourceBusinessWire, Aug 2025 (opens in a new tab)Tempus Announces the Acquisition of Paige Low confidenceInferred, implied or estimated, so read it as rough. | ||||
| Books, documents and expert work6 prices, per title, per book, per question, per million words, per yearExplore documents | ||||
| Nonfiction backlist licenseSplit 50/50 with authors who opt in | $5,000per titlefor 3 years | High confidenceeWeek (opens in a new tab) Dec 2024 | ||
What this price coversHarperCollins' three-year license of nonfiction backlist titles for AI training, split 50/50 with authors who opt in. What moves it
SourceeWeek, Dec 2024 (opens in a new tab)HarperCollins Book Catalog To Train Microsoft AI Models For The Next 3 Years High confidenceA completed deal or settlement disclosed with its price. | ||||
| Book settlement benchmarkA court settlement, not a license price | About $3,000per book | High confidenceTechCrunch (opens in a new tab) Jul 2026 | ||
What this price coversWhat a $1.5B settlement with authors, approved in July 2026, worked out to per book. It is a court settlement, listed as a benchmark, not a license price. What moves it
SourceTechCrunch, Jul 2026 (opens in a new tab)Anthropic's landmark $1.5B copyright settlement is approved High confidenceA completed deal or settlement disclosed with its price. | ||||
| Expert exam questionsBenchmark prizes for the hardest questions | $500 to $5,000per question | Midpoint of a sourced range: $1,580 per question. | Medium confidenceScale AI (opens in a new tab) Sep 2024 | |
What this price coversPrizes a benchmark program paid for the hardest expert-written questions: $5,000 each for the top 50 and $500 for the next 500. Midpoint of a sourced range: $1,580 per question. What moves it
SourceScale AI, Sep 2024 (opens in a new tab)Submit Your Toughest Questions for Humanity's Last Exam Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Translation memoriesSpecialist and rare pairs hold value | EUR 1,500 to 2,500per million wordsthen cut 97%+ | Midpoint of a sourced range: $2,200 per million words, in US dollars at $1.1355 per euro. | Low confidenceMultiLingual (opens in a new tab) Apr 2024 | |
What this price coversList prices for translation memories on a data marketplace, before it cut its prices by more than 97%. Specialist and rare language pairs hold value. Midpoint of a sourced range: $2,200 per million words, in US dollars at $1.1355 per euro. What moves it
SourceMultiLingual, Apr 2024 (opens in a new tab)TAUS Data Sale to Boost Multilingual LLMs Low confidenceInferred, implied or estimated, so read it as rough. | ||||
| News licensing, large publishersAverage of 34 disclosed deals | About $24Mper yearaverage | Medium confidenceMedia and the Machine (opens in a new tab) Jan 2025 | ||
What this price coversThe average yearly value of 34 disclosed AI licensing deals with large news publishers. Smaller publishers use pooled and usage-based programs. What moves it
SourceMedia and the Machine, Jan 2025 (opens in a new tab)5 Takeaways from the AI Content Licensing Deals I'm Tracking (size of avg. deal , market, etc.) Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms. | ||||
| Community Q&A (forum archive)One platform's license with Google | About $60Mper year | High confidenceColumbia Journalism Review (opens in a new tab) Oct 2025 | ||
What this price coversOne large community platform's yearly license with Google for its posts and answers. What moves it
SourceColumbia Journalism Review, Oct 2025 (opens in a new tab)Reddit Is Winning the AI Game High confidenceA completed deal or settlement disclosed with its price. | ||||
| Robotics and sensors2 prices, per hour, per datasetExplore robotics | ||||
| Robot teleoperation logsEstimates; force and touch data raise it | $50 to $200per hour | Midpoint of a sourced range: $100 per hour. | Low confidenceDreamVu (opens in a new tab) Aug 2026 | |
What this price coversLogs of robots driven by human operators, with their camera and sensor streams, priced per hour of robot data. These are industry estimates, so the row is marked Low. Midpoint of a sourced range: $100 per hour. What moves it
SourceDreamVu, Aug 2026 (opens in a new tab)Robot Training Data Companies: The 2026 Landscape Low confidenceInferred, implied or estimated, so read it as rough. | ||||
| Industrial sensor dataNo public rate card; labeled faults matter most | Priced after reviewper dataset | No public per-unit price | Low confidencearXiv (opens in a new tab) Jun 2026 | |
What this price coversMachine and factory sensor logs. There is no public rate card, so each dataset is priced after review. What moves it
SourcearXiv, Jun 2026 (opens in a new tab)FactoryNet: A Large-Scale Dataset toward Industrial Time-Series Foundation Models Low confidenceInferred, implied or estimated, so read it as rough. | ||||
How the index works
- What a row is
- One kind of data at one price, in one unit, from one dated source: a disclosed deal or settlement, a winning bid, a pre-sale estimate, a reported range or program term or, where nothing is public, our own estimate. Sources are press reports, company disclosures, published buyer programs and industry research.
- Units
- Each price is per its own unit: per repo, per company, per hour and so on. Rows compare directly only when they share a unit, so pick one in the unit menu and the axis rescales to it.
- Confidence
- High confidenceA completed deal or settlement disclosed with its price.
- Medium confidenceReported ranges, averages, winning bids, pre-sale estimates, listings or program terms.
- Low confidenceInferred, implied or estimated, so read it as rough.
- Bars and midpoints
- A bar spans each range on a log scale. The dot is the midpoint of a sourced range: the geometric mean of its two ends, which is the visual center of a bar on a log axis, or the reported average where a row gives one. A single reported price is a dot, and a bar with an open top fades out.
- Currency
- US dollars unless a row says otherwise. Rows priced in euros keep their currency and are plotted at $1.1355 per euro, the reference rate for 29 Sep 2026. European Central Bank, (opens in a new tab)
- Updates
- A new edition each quarter; the next is due in January 2027. This edition was published on 30 Sep 2026 and updated on 2 Oct 2026.
- Not an offer
- These are market reference points, not offers. Your data is priced on its own scarcity, structure and rights, which is what the appraisal measures, and every range it shows reads: Indicative estimate, not an offer. Final offers depend on the systems you connect, data size and quality, access terms and due diligence.
- Corrections
- Found a better source or a newer deal? Email hello@rightmark.ai and we review it for the next edition.
Edition notes
First edition, 30 Sep 2026
New4
This first edition covers 23 prices in 7 data families, from $0.02 per stock image to about $60M a year for a community forum archive. Reuters via Rappler, (opens in a new tab); Columbia Journalism Review, (opens in a new tab)
Private code repositories enter at $150 to $1,500 per repo per sale, from pre-sale estimates in code licensing programs. Rightmark,
Operating company archives enter at $185,000 to $1M+ per company, from published buyer programs. Rightmark,
Robot teleoperation logs enter at $50 to $200 per hour from industry estimates, so the row carries Low confidence. DreamVu, (opens in a new tab)
Context4
A $10M winning bid in a bankruptcy auction for one company archive still needs court approval, and a rival bidder later offered more, so the row carries Medium confidence. TIME, (opens in a new tab); Fortune, (opens in a new tab)
The book settlement benchmark of about $3,000 per book comes from a court settlement with authors, not from a license. TechCrunch, (opens in a new tab)
Translation memories were listed at EUR 1,500 to 2,500 per million words, then cut 97%+, so the row carries Low confidence. MultiLingual, (opens in a new tab)
The premium repository range is our own estimate, because no buyer publishes a per-repo price for task-grade code. Rightmark,
Watching4
An education company announced a move into AI model training built on its subject matter experts and proprietary data; the index adds a row when a per-item price is reported. Chegg, (opens in a new tab)
A large news publisher reported a yearly AI licensing deal above the $24M average of the disclosed news deals in this edition. Yahoo Finance, (opens in a new tab)
A large question-and-answer site now licenses its content to AI developers for a flat fee; the index adds it when a price is reported. TechCrunch, (opens in a new tab)
Lawsuits have been filed against several large tech companies over voiceprints used to train AI; consent that names AI training is the safer basis for a voice data deal. Biometric Update, (opens in a new tab)
Questions about the index
How much do AI companies pay for data?
It depends on the type and unit. In Q3 2026: $150 to $1,500 per private code repository per sale at standing rates, $1 to $4 per minute of unused video, $50 to $150 per hour of speech, up to $5,000 per book title, and $10,000 to $1M or more per company archive.
Sources: Rightmark, ; The Decoder, (opens in a new tab); SpeechData.ai, (opens in a new tab); eWeek, (opens in a new tab); Rightmark,
Where do the prices come from?
From press reports, company disclosures, published buyer programs, pre-sale estimates and industry research. Each row records its unit, date and source. Where no public price exists we say so, or show our own estimate marked Low confidence.
How often is the Price Index updated?
Every quarter. The current edition is Q3 2026, published 30 Sep 2026, and the next is due in January 2027. Past editions stay online at their own addresses, so citations do not break.
What do the confidence levels mean?
High means a completed deal or settlement disclosed with its price. Medium means reported ranges, winning bids, pre-sale estimates or program terms. Low means the price is inferred or estimated and should be read as rough. Revenue-share rows have no public per-unit price.
How much is an hour of speech data worth?
Ready-made speech datasets are reported at $50 to $100 per hour, and low-resource languages at $60 to $150. Bulk call-center audio lists at $6.30 to $7 per hour.
Sources: SpeechData.ai, (opens in a new tab); Datarade via Internet Archive, (opens in a new tab)
How much do AI companies pay for books?
Reported book prices run $3,000 to $5,000 per title: about $3,000 per book from a court settlement with authors, a benchmark rather than a license price, and $5,000 per title for 3 years for a nonfiction backlist license shared with authors who opt in.
Sources: eWeek, (opens in a new tab); TechCrunch, (opens in a new tab)
Can I cite or reuse the Price Index?
Yes. Cite it as Rightmark Price Index, Q3 2026, Rightmark, with a link. The tables are free to reuse with attribution under CC BY 4.0, and a CSV download sits on the page.
Cite this edition
Rightmark Price Index, Q3 2026. Rightmark. Published 30 Sep 2026, updated 2 Oct 2026. https://rightmark.ai/price-index/
Free to reuse with attribution under CC BY 4.0 (opens in a new tab). The CSV holds every row with its unit, confidence, source and source link.
Put your data on the board
Answer a few questions to see an indicative range for your data, marked on this board where the index has a matching price.