worksheet
RL environment budget worksheet
Enter how many tasks you need, how they are graded and which rights you want. The worksheet multiplies the only published per-task range (Epoch AI's) by your count and lists the questions that decide where a real quote lands.
basis: Epoch AI, 12 Jan 2026
How much should we budget for an RL environment?
full list
What should we ask every RL environment vendor?
The worksheet shows the questions for your choices. Here is the full set.
For any project
- Do we receive the environment code and tasks, or hosted access only?
- Who owns the trajectories our training runs produce?
- What does a paid pilot of a small task set cost, and what would make you reprice the full order?
Automatically checked tasks
- How many tasks can reuse one sandbox image, and what does a new image add?
- How are graders tested against wrong solutions before delivery?
- How do you find and remove near-duplicate tasks?
Multi-step tasks
- How many steps does a typical task take, and how is partial credit scored?
- Which tools does the agent get, and are they documented per task?
- What share of tasks did your own reviewers fail to solve, and what happened to them?
Rubric or expert-graded tasks
- Who writes the rubrics, and what is their background in the domain?
- Is grading done by a model judge, a person, or both, and how do you check agreement?
- What does a re-grade cost if we change the rubric later?
Rights
- Which other buyers can license the same tasks, and is that list disclosed?
- How long does exclusivity last, and does it cover the graders and the sandbox code too?
- What is the price for each option, shared and exclusive, on the same task set?
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).
4 questions
Questions about the worksheet
- Where does the worksheet's range come from?
- From Epoch AI's FAQ on RL environments (12 Jan 2026): $200 to $2,000 per task in most cases, and 4 to 5 times that for exclusive deals. Both are interview estimates.
- Why does complexity not change the number?
- No source publishes cost by complexity, so the worksheet does not guess. Complexity changes the questions you should ask instead, because it changes what drives the quote.
- Is the result a quote?
- No. It is a planning bracket built on published estimates. Vendors quote per project, and actual prices can land well outside it.
- Does the result include training compute?
- No. It covers building tasks and graders only. Running training is a separate cost.
see also
Related pages
- Task costs: the published figures in full
- Companies: who to ask
project brief3 parts
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.