RL environment pricing: what labs pay in 2026

The short answer

Almost every RL environment price in public traces back to three sources: Epoch AI's FAQ, based on 18 interviews with founders and lab researchers (January 12, 2026), SemiAnalysis (January 6, 2026), and The Information (September 2025). Treat every number below as a reported range from a private market, not a list price. Epoch itself says the market will look very different within a year.

Price points, with sources

What drives the cost

Fidelity of the simulated product

The gap between $20,000 for a website replica and $300,000 for a Slack-class replica is mostly fidelity: how much of the real product's behavior, data model and edge cases the clone reproduces. More is not always better. Prime Intellect's Will Brown advises cutting simulator complexity that is "overkill," for example using an in-memory mock instead of a full database (Sequoia podcast). Pay for fidelity only where the skill you are training depends on it.

Domain expertise

One of Epoch's interviewees said "domain knowledge and expert level prompting is more important than ML skills for creating tasks." A finance or law task needs someone who does that work to write it and to check the grader.

Verifier hardening

"It takes many many iterations to check against reward hacking," a founder told Epoch. Each round means running a strong model against the task, reading the trajectories that scored high, and patching the grader. Cheap tasks often skip this step. See our evaluation checklist for how to test it.

Difficulty calibration

A task must sit in a pass-rate band on your model. Calibrating takes rollouts, which cost compute, and sometimes a blind human solve to prove a 0% task is feasible. Ask whether the vendor calibrated against your checkpoint or against a public model, and who pays for those rollouts.

Exclusivity

Most vendors "close source their environments and serve them in exclusive contracts to the labs" (SemiAnalysis). Exclusive rights cost about 4 to 5 times as much because the vendor cannot resell the work. Labs pay it to keep training priorities private.

Quality at scale

"Maintaining quality while scaling is the number one bottleneck," per Epoch's sources. Founders said revenue growth is limited by how fast they can create quality tasks. Our reading: while supply is the constraint, prices for well-built tasks will stay high.

Hosting and infrastructure

Environments mostly run on CPU servers next to GPU training (SemiAnalysis, June 2025). They need many parallel rollouts, fault tolerance and fast resets. We found no public source that prices hosting separately, so ask whether it is included.

Worked budget example

This is arithmetic on the cited figures, not a quote. The scenario and its quantities are ours. The unit prices are the reported ones above.

Suppose a neolab plans one quarter of agent training with three purchases, all non-exclusive:

Three things fall out of the arithmetic.

• The non-exclusive total lands in the "six to seven figures per quarter" range Epoch and Sapphire report, which is a sanity check on the scenario.
• At the low end of task prices, compute for the coding tasks ($960,000) is about 12 times their purchase price ($80,000). Even at the high end it is more than the tasks themselves cost. The replicas host many tasks each and add their own compute on top.
• Suppose 100 of the 400 tasks turn out to sit at 0% or 100% on your model. Under group-relative methods like GRPO, those tasks produce zero advantage and no gradient (Shao et al.). At Mechanize's estimate that is about $240,000 of compute with no learning signal, plus $20,000 to $200,000 paid for the tasks.

The lesson from the third line: the cheapest way to cut cost is to reject bad tasks before training, not to negotiate the per-task price.

How pass rates connect to price

Epoch reports buyers want at least one success in 64 to 128 rollouts, which is a minimum pass rate of roughly 2 to 3%. The probabilities show why the two framings line up:

Computed as 1 − (1 − p)ⁿ. Put a pass-rate floor in the contract, measured on your checkpoint, with a rule for replacing tasks that fall below it. Epoch's sources also discard tasks near 70%, so a ceiling belongs there too.

How labs buy

• Spec, build, review. A lab research team writes a task spec, the vendor builds, and the lab reviews and stress-tests. Some labs route purchases through data-ops teams; others let researchers buy directly (Sapphire, April 2026).
• Sample first. Labs typically test a sample environment, including probing the verifier for reward hacks, before signing, according to a vendor guide from Troveo.
• Replicas are bought once. UI gyms are a one-time purchase reused across later runs (SemiAnalysis). Tasks lose value as the policy saturates them.

Market context

The Information reported in September 2025 that Anthropic leaders discussed spending over $1 billion on RL environments over the following year, as cited by TechCrunch and Epoch. That was a reported discussion, not a confirmed budget. Epoch publishes no total market size. It does note that compute will far exceed environment spend: OpenAI's 2026 R&D compute is projected at about $19 billion. For comparison, Wing VC sizes traditional data labeling at about $5 billion, growing over 50% a year (January 2026). That figure covers labeling, not environments.

What we could not find

• Per-rollout, per-seat or revenue-share pricing. We found no reputable public source for any of these.
• Typical build times per environment, or contract lengths beyond per-quarter pricing.
• Standard IP and data-rights terms. Epoch does not cover them, and we found no primary public source.

If you have a source for any of these, send it to us and we will add it. For background, see what RL environments are and what an RL gym is. For free alternatives, see open-source RL environments. For terms, see the glossary.