Where does the figure come from?
There is one public source worth quoting on this: Epoch AI’s “An FAQ on reinforcement learning environments”, by Denain and Barber, published 12 Jan 2026. Epoch talked to people who build and buy environments and asked what a task costs.
The answer it reports is “$200 to $2000 mostly. $20k per task would be rare but possible”. That is for one task including its grader, the code or rubric that decides whether an attempt passed.
Two caveats travel with that range. It comes from interviews, not from invoices or signed contracts. And it describes the market at the start of 2026. Neither makes it useless; both mean it is a starting point for a budget, not a number to hold a vendor to.
What is a “task”, exactly?
The word covers more than an instruction. A task in an RL environment is:
- a starting state (a repository at a commit, an app with seeded data, a desktop with files open);
- an instruction the agent receives;
- a grader that checks the end state;
- often a reference solution proving the task can be done.
A cheap task reuses an existing sandbox, has an obvious check and takes a writer an hour. An expensive one needs a new app replica, a domain expert to write it, and several rounds of testing the grader against wrong answers.
Why is the range so wide?
A tenfold spread between $200 and $2,000, with outliers ten times higher again, reflects how different tasks are. The main drivers:
| Driver | Cheaper end | Costlier end |
|---|---|---|
| Grading | Automatic test or exact answer | Rubric, model judge or expert review |
| Writer | Generalist engineer | Domain specialist (law, finance operations, a niche codebase) |
| Sandbox | Shared image reused across tasks | A new replica built for a few tasks |
| Length | A few steps | Long, multi-tool workflows |
| Review | One pass | Adversarial testing and repair |
None of the companies on our list publishes where its work falls in this range. What some do publish is scale. Scale AI’s RL environments page lists Enterprise Atlas, four enterprise environments with more than 2,000 tasks, and computer-use sets of 1,000 and 370 tasks. Appen’s blog counts 500 coding tasks, every one with a checker that runs by itself. Those counts show the order of size vendors work at, without telling you the price.
How much do exclusive rights add?
Epoch AI’s FAQ says exclusive arrangements cost about 4 to 5 times what shared ones do. On the mid-point of the per-task range, that turns a $1,100 shared task into something like $4,400 to $5,500 exclusive.
Whether that premium is worth it depends on what the tasks teach. For broad skills that every lab is training, owning the tasks gives little lasting edge. For a workflow tied to your own product, it may be the point of the purchase.
What does that mean for a whole project?
The arithmetic is simple, and the table on our task costs page does it for several task counts. As an example on Epoch’s range:
- 500 shared tasks: $100,000 to $1,000,000.
- 500 exclusive tasks: $400,000 to $5,000,000.
- 2,000 shared tasks: $400,000 to $4,000,000.
Each line is multiplication on an interview-based range. It is not an offer, and actual quotes can sit far outside it in either direction.
What is left out of the per-task figure?
Building tasks is one cost among several. Budget separately for:
- Training compute to run the agent through the environment, often many times per task.
- Hosting, if the vendor runs the sandbox for you instead of handing over code.
- Iteration, because the first training runs will expose tasks that are broken or too easy, and fixing them costs time.
- Your own team’s review, which no contract removes.
Why do so few vendors publish prices?
Every company on our list works on custom quotes. Part of the reason is the spread above: a single rate would be wrong for most buyers. Part is that the work is sold mostly to a small number of large labs, each negotiating its own terms on volume, rights and delivery. A published rate card would also tell competitors where to undercut. Until that changes, interview-based estimates like Epoch’s are the closest thing to a public benchmark.
How should a team use this when talking to vendors?
Use the range to sanity-check, not to negotiate line by line. If a quote lands far below $200 per task, ask what is being skipped, usually grader testing. If it lands far above $2,000, ask what makes the tasks unusual: a new sandbox, rare expertise or exclusivity. Both conversations tell you more about the vendor than the number does.
When you compare RL environment companies, ask each for a priced pilot of a small task set first. The pilot shows the true cost per task for your domain, and that beats any published estimate.
A good task costs what it takes to make the grader right; everything else is detail.