What does one RL environment task cost?
Mostly $200 to $2,000 for one task with its grader, according to interviews by Epoch AI published 12 Jan 2026; $20,000 for a task is rare but possible. No vendor publishes a rate card, so these are market estimates, not prices.
Last verified 7 November 2026
What has been published on RL task cost?
Epoch AI, a research group that tracks AI inputs, asked people who build and buy environments what a task costs. Its FAQ gives the answer as "$200 to $2000 mostly. $20k per task would be rare but possible". On rights, it puts exclusive arrangements at about 4 to 5 times the cost of shared ones.
Both figures come from interviews, not invoices or contracts, and Epoch says so. They are still the only public numbers we found, and they agree with what the companies themselves show: every one prices on a custom quote.
What would a set of tasks cost on that basis?
Multiplying Epoch's range by a task count gives a rough bracket. It is arithmetic on an interview-based range, not an offer; actual quotes may be well below or above it.
| Tasks | Non-exclusive, at $200 to $2,000 each | Exclusive, at 4 to 5 times that |
|---|---|---|
| 100 | $20,000 to $200,000 | $80,000 to $1,000,000 |
| 500 | $100,000 to $1,000,000 | $400,000 to $5,000,000 |
| 1,000 | $200,000 to $2,000,000 | $800,000 to $10,000,000 |
| 5,000 | $1,000,000 to $10,000,000 | $4,000,000 to $50,000,000 |
The environment budget worksheet does the same sum for your own count and adds the questions to put to vendors.
What makes one task cost more than another?
How it is graded
A task checked by running a test is cheaper than one that needs a written rubric, a model judge or a human reviewer.
Expert time
Tasks in law, finance operations or a specialist codebase need writers who know the field, and their hours dominate the cost.
The sandbox
Tasks that reuse an existing environment cost less than ones that need a new app replica or a fresh container image.
Checking
Testing each grader against wrong answers, and fixing tasks an agent can game, adds review time per task.
Rights
Exclusive tasks cost several times shared ones, on Epoch's figures.
Where these facts come from (1)
- Epoch AI, 12 January 2026: "An FAQ on reinforcement learning environments" (Denain and Barber) (page checked 7 November 2026).
Questions about task cost
- Is $200 to $2,000 a price I can expect to be quoted?
- No. It is the range Epoch AI heard in interviews with people in the market, not a published rate. Treat it as a basis for a budget conversation; your quote depends on the vendor, the domain and the terms.
- Why would one task cost $20,000?
- Epoch AI calls that rare but possible. Costs of that order would fit tasks that need long expert work to write and check, or a custom sandbox built for a single task.
- Does the per-task figure include compute?
- Epoch's range is for building a task with its grader. Running training on it is a separate cost, paid to your compute provider or the vendor if it hosts the runs.
- Do any vendors publish rate cards?
- Not for RL environments. None of the 14 companies on our companies page publishes one (checked 7 November 2026).
- How much more does exclusivity add?
- On Epoch AI's figures, a multiple of 4 to 5 over the shared price of the same work.
Related pages
- Environment budget: the worksheet
- Companies: who builds environments
- What one RL training task costs to build: the longer piece
Get a quote
Tell us the company ZIP code, a contact, and an outline of the environments you have in mind. We compare it with what each builder publishes and send you the shortlist by name.
How the site earns: we make introductions, and a builder you later contract with may pay us for yours.
Received. The next three things:
- Your outline is set against what each RL environment builder says it makes and the terms it publishes.
- A shortlist arrives by email: names, links, and a sentence on the fit for each.
- No builder hears about you until you answer yes. There is no mailing list to join.