You build graded tasks and environments. Labs buy them ready-made, and pay for graders that cannot be fooled.
Teams and experts who build training environments, graded task sets and private evaluations: in software, CAD and PCB design, chip design, formal proofs, science and professional work.
What companies like yours hold
What you typically have, where it sits, and whether the valuation can measure or value it. Buyers price what exists together, over how many years, and how well it links.
- Graded task setsWhere it sits, and what the evidence showsTasks with an automatic grader each, in containers that run anywhere. The engineering set a lab buyer chose on 12 October held 12K+ containerised tasks, all graded automatically.How the valuation reads itValue a task setpriced on its own, per task
- Engineering design tasksWhere it sits, and what the evidence showsCAD and PCB design (3,400 tasks in that set) and chip design with testbenches and coverage (9,400).How the valuation reads itValue a task setpriced on its own, per task
- Formal proofsWhere it sits, and what the evidence showsLean proof tasks checked by the proof assistant (1,500 in that set).How the valuation reads itValue a task setpriced on its own, per task
- Research and writing tasks scored on rubricsWhere it sits, and what the evidence showsHypotheses, experiment designs, grant sections and responses to peer review. The science environment chosen the same day had 400+ evaluation tasks across 7 tracks.How the valuation reads itValue a task setpriced on its own, per task
- Simulated apps and workplacesWhere it sits, and what the evidence showsWorking replicas of software for an agent to practise in, with seeded data and checks.How the valuation reads itDescribe it in your offer
- Private evaluation setsWhere it sits, and what the evidence showsQuestions or tasks never published, with expert solutions.How the valuation reads itValue a task setpriced on its own, per task
- Build and run historyWhere it sits, and what the evidence showsThe repositories that build each environment, and logs of model attempts and grader results.How the valuation reads itDescribe it in your offer
What buyers say they want from companies like you
In their own words, from their own pages or from reporting on what they bought. Each opens its source.
“Math and coding tasks came first, but we're now seeing significant growth in enterprise workflows: tasks like navigating Salesforce, filing reports, or manipulating spreadsheets.”
Epoch AI interviewsAs reported, 12 Jan 2026 “Contract sizes are often six to seven figures per quarter.”
Epoch AI interviewsAs reported, 12 Jan 2026 “Opportunities are usually priced per task, at around $500 to $1,500 a task, with hundreds to thousands of tasks per request.”
HUD DataVendorIts own page, 12 Oct 2026 “Scale AI wrote in February that nearly half of its new data-training projects involve reinforcement-learning environments.”
Scale AI, per FortuneAs reported, 1 Oct 2026 “private, uncontaminated eval sets because public benchmarks have leaked into pretraining data.”
Labs, per a practitioner (Gokul Rajaram)As reported, 11 Oct 2026
What raises the value
- Sound graders: "Soundness matters most: high reward must mean the task was actually solved, not hacked" (Epoch AI interviews).
- Difficulty set so current models pass sometimes, not always.
- A professional domain where the builder has real expertise: engineering, chip design, proofs, science, finance, law.
- Scarce source material: chip designers "guard their IP closely", so little open data exists to train on (TechCrunch, 1 Apr 2026).
- Tasks never published anywhere, so no model has seen them.
- Exclusivity, when the buyer pays for it: "exclusive deals are roughly 4-5x more expensive than non-exclusive ones" (Epoch AI interviews).
- Containers that build and run cleanly, with documentation a lab's team can use without the builder.
What lowers it
- Plain maths problems and static benchmark-style tasks: "math tasks are easy to create but don't transfer as well to other capabilities" (Epoch AI interviews).
- Quickly built replica websites: "what everyone does is vibe code a buggy website which isn't useful" (Epoch AI interviews).
- Tasks current models already pass. In one practitioner's words, "The product depreciates by succeeding."
- Graders that reward shortcuts or can be gamed.
- Tasks that leaked online or were derived from public benchmarks.
A line that quotes a buyer names it. The others are our checklist, drawn from the factors buyers state in general: no buyer publishes how much each one moves a price. How often each factor matters across 156 kinds of data.
The rights to settle first
These stop a sale more often than price does. Each can be settled before any buyer sees a record.
- You must own the environment and everything inside it. Tasks derived from someone else's proprietary material carry that material's restrictions.
- Chip-design tasks: foundry design-kit and EDA tool licence terms, third-party IP cores and export control.
- Tasks written by contractors or experts need an assignment of rights to you, and their contracts must allow resale.
- Proof and science tasks built on published papers or problem sets carry the source's licence.
- Selling a private evaluation set exclusively ends your own use of it; selling it to several buyers lowers what each pays.
A recent purchase that fits
One of five lots a lab buyer went for on 12 Oct 2026, wanting to close in under a week. No price was given.
- Seller
- Builder of engineering training environments
- People
- Not stated
- History
- Not stated
- Size
- 12K+ containerised tasks
- What was inside
- 3,400 CAD and PCB design tasks
- 9,400 chip-design tasks
- 1,500 Lean proof tasks
- All graded automatically
Source: A buyer-side referral network's announcement to its members, 12 October 2026 (reported to us).
What is already on offer from companies like you
Read this morning from the brokers' and marketplaces' own catalogues. A buyer compares your records with these; settled rights, more years and more linked systems are what set one apart.
lots on offer like yours
of history in the middle of them
say the rights are settled
- Longitudinal company workforce panel, 1950-202070 years · Grepped marketplace
- Texas construction cost data: job costs, bids, change orders, quotes and takeoffs23 years · Grepped marketplace
- Financial models and accounting workbooks (Excel)21 years · Grepped marketplace
- Consented European consumer panel: search, browsing, video, social, shopping, AI-chat exports and mobile computer-use20 years · Grepped marketplace
- Tax-preparation case corpus11 years · Grepped marketplace
- Chest and general X-rays with radiologist reports9 years · Grepped marketplace
- Merchant-built fashion lookbooks: 700K+ outfit bundles8 years · Grepped marketplace
- Tech-candidate profiles with interview outcomes (India, US)8 years · Grepped marketplace
What holders like you ask
- Who buys finished environments?
- Labs buy mostly through vendors and marketplaces rather than from builders: of 15 labs we checked, only Google runs a self-serve route for outside sellers. DataVendor (HUD) lists environments and task sets from builders for lab research teams, Prime Intellect pays bounties for environments it releases open source, and on 12 October one lab buyer chose two environment lots, an engineering task set and a science research environment.
- What does a graded task sell for?
- HUD DataVendor says opportunities are "usually priced per task, at around $500 to $1,500 a task, with hundreds to thousands of tasks per request". The valuation prices a task set from what the market's records say one task fetches in its field, times how many tasks it holds; where the middle lands follows what task buyers ask for: graded automatically, running as an environment, never published. Exclusive deals run roughly 4 to 5 times a non-exclusive one (Epoch AI interviews). No price was given for the two environment lots chosen on 12 October.
- Do maths and coding tasks still sell?
- Plain maths problems sell less well: "math tasks are easy to create but don't transfer as well to other capabilities" (Epoch AI interviews). Formal proofs differ in one way buyers care about: a proof assistant checks them, so the grade cannot be gamed. The engineering lot chosen on 12 October held 1,500 Lean proof tasks beside 12,800 CAD, PCB and chip-design tasks.
- How long does an environment stay valuable?
- Not long. Practitioners put an environment's useful life at about six months, because models learn to pass it: "The product depreciates by succeeding." Sell while current models still fail some of the time, and keep building.
- Should we sell exclusively?
- It is a price, not a default. Exclusive deals run roughly 4 to 5 times a non-exclusive one (Epoch AI interviews), but an exclusive sale ends your ability to sell the same set again, and an environment's useful life is short. A sale with several buyers shows what each pays for each kind of licence before you choose.
Other kinds of holder
In this topic
What sells
- Start hereWhich data do AI labs want, and who holds it?Each kind of data the market prices, the businesses that hold it, what labs are asking for now and what they turn away.
- GuideWhat Are RL Environments? Who Builds, Buys and Prices ThemAn RL environment is a sandbox where an AI agent attempts tasks with tools and a verifier scores each try. Who builds them, who buys them and what they cost.
- GuideHow to Sell Videos to AI Companies: Prices, Buyers and RightsAI companies license video by the hour or minute. What buyers pay, which footage sells, and the rights you need in every frame before you sell a library.
- PageWhere data comes fromWho holds each kind of data, how much exists and how it reaches buyers.
- RegisterCompanies closingCompanies closing, sold or gone dark, and the records they leave.