What RL Environments Cost in 2026

No vendor in this market publishes a price list. The numbers buyers pass around come from two public sources, Epoch AI's survey of the RL environments industry and TechCrunch's reporting on lab spending, so this page starts with those numbers and then explains what actually moves them.

How much do RL environments cost?

Public reporting puts RL environment costs at roughly $200 to $2,000 per task for most task-based work, around $20,000 for a high-fidelity website replica, up to about $300,000 for a complex application clone, and six to seven figures per quarter for frontier-lab contracts. Breaking that down by source:

Per task: Epoch AI's survey of environment builders reports most tasks priced at $200 to $2,000, with complex software engineering tasks reaching $20,000 each in rare cases.
Per replica: the same survey reports website replicas, sometimes called UI gyms, at around $20,000 each, while a clone of a complex application like Slack can run around $300,000 (Epoch AI).
Per contract: lab contracts commonly land in the $300,000 to $500,000 per quarter range, with the largest at seven figures per quarter or more (Epoch AI).
Exclusivity: exclusive deals run roughly 4 to 5 times the price of non-exclusive ones (Epoch AI).
Total market signal: TechCrunch reported in September 2025 that leaders at Anthropic had discussed spending more than $1 billion on RL environments over the following year, a figure Epoch AI also cites.

Those are frontier-lab numbers. If you are a company buying an environment for one of your own business processes rather than selling task batches to a lab, the shape of the cost is different, which is what the rest of this page covers.

What drives the cost of an RL environment?

Five things drive the cost: task complexity, replica fidelity, verification depth, expert time for annotations, and maintenance as the underlying process changes.

Task complexity. The spread in Epoch's per-task numbers, $200 at the bottom and $20,000 at the top, is almost entirely a complexity spread. A task that takes an expert an hour to specify and check costs a fraction of one that takes days.
Fidelity of the replica. A form with ten fields is cheap to reproduce. A faithful clone of a large application is not, which is the gap between the ~$20,000 website replica and the ~$300,000 Slack-class clone in Epoch's reporting.
Verification depth. A grader that string-matches an answer is cheap. A grader that is robust to reward hacking is not. Epoch's interviewees describe grader hardening as iterative work that rarely succeeds on the first pass. How checks get built, and why some are cheap code and others need trained judges, is covered in our RLVR guide.
Expert time for annotations. Judged checks need operators to annotate pass/fail examples, and operator hours are often the scarcest input in the build.
Maintenance. Business processes change. When the ERP gets upgraded or the approval chain changes, the environment and its checks have to change too, which is why ongoing cost exists at all.

How are RL environments priced?

Four pricing models exist in the market today: per-task, per-environment build, subscription platforms, and consultative build-and-retainer.

Per-task fits labs buying training data in volume. You pay for a batch of tasks with graders, at the $200 to $2,000 per task rates above.
Per-environment build fits one-off replicas, priced like the $20,000 UI gyms or the larger app clones.
Subscription and platform models charge for access to hosted environments and tooling rather than for a specific build.
Consultative build-and-retainer is what Pebble does: scoped work packages covering discovery, implementation, and review, plus a maintenance retainer once the environment is live. The engagement runs scope (a call), discover (about a week inside your systems, and we keep the step map and none of your data), build, then run with a handoff and retainer.

To be direct about it: Pebble does not publish list prices, and neither does anyone else in this market. The proposal comes out of a scoping call, because the price is mostly a function of how many steps your process has, how many of the checks are computable versus judged, and how much operator annotation time the judged checks need. None of that is knowable from a pricing page. A comparison of who builds what is in our rundown of RL environment companies.

When does owning the environment beat renting frontier API calls?

Owning wins when run volume is high enough that per-run inference cost dominates, and when you cannot afford to have the model behind your process repriced or deprecated by someone else. Pebble's engagement ladder makes this a staged decision rather than a bet. Stage one, RUN, is working software on one process using a frontier model. Stage two, PROVE, adds a full audit trail so reliability becomes a percentage read off run history rather than a claim. Stage three, OWN, runs an open-source model through the environment at volume, with passing runs weighing more in training, until you have a model you own that no one can reprice or deprecate out from under you.

We will not put invented numbers on the crossover point, because it depends on your run volume and your process. The qualitative math is straightforward: self-hosted batch inference on an open-source model costs a fraction of frontier API pricing per run, and the environment is what makes the open-source model good enough to trust. The audit trail from PROVE is what tells you whether you have reached that point, which is the reason the ladder is ordered the way it is.

Why is there no public price list for RL environments?

Because no two environments cost the same amount to build, and the market's demand side has so far been a handful of labs negotiating private contracts. Epoch's reported range spans two orders of magnitude on per-task pricing alone, and contract structures vary further with exclusivity, which by itself moves price by roughly 4 to 5 times. A list price would be wrong for almost every buyer.

Do you need a frontier-lab budget to buy an environment?

No. The seven-figure quarterly contracts in Epoch's reporting are labs buying environments in bulk to train general-purpose models. A single business process, mapped into steps with a check per step (terms defined in the glossary), is a much smaller build, and it starts as working software on one process rather than a training-data contract. Which vendors price for which buyer is laid out in the RL environments market map.

What is the first step?

A scoping call with the founder who builds the environments: [email protected]. Bring one process. If an environment is the wrong tool for the job, we say so on the call.