Building RL environments in-house gives you full control over tasks, data and verifiers, and frontier labs do much of their environment work this way. The cost is engineering time — at Mechanize, which builds coding environments, one engineer spends about a week on each task — plus the ongoing work of keeping verifiers honest as models improve. Build for proprietary workflows and anything you must own outright; buy for common app categories, speed, or an independently built and tested verifier.
How does building compare with buying at a glance?
| Factor | Build in-house | Buy from a builder such as Sandworld |
|---|---|---|
| Control over tasks and verifiers | Full | Shared; inspect and rerun the vendor’s QA |
| Time to first tasks | Weeks to months of engineering | Immediate for existing twins; a pilot for a custom one |
| Cost profile | Engineering time, infrastructure and ongoing QA | Per-task or per-environment pricing |
| Intellectual property | You own everything | Depends on the license; exclusivity costs more |
| Independent check on graders | Only if you staff a separate team | A second team built and attacked the verifier |
| Expert demonstrations | Recruit and manage your own experts | The vendor’s program (Sandworld’s is just starting) |
What does building an RL environment in-house involve?
- A backend that holds realistic state, with seeded variation and a simulated clock.
- Typed tools with argument validation, and a policy the agent must follow.
- Tasks phrased as real requests, including traps where the right answer is to refuse.
- Verifiers that recompute answers per seed and check for collateral damage.
- QA: reference solutions on many seeds, plus do-nothing, random, truncated and noisy probes.
- Calibration runs per model, so tasks stay learnable, and retirement as they saturate.
- Isolation and operations. Sandboxes have to be locked down — Prime Intellect disclosed an escape from “offline” sandboxes — and resources pinned: Anthropic found that resource configuration alone moved Terminal-Bench 2.0 scores by 6 points.
The verifier is usually the hardest part. A 2026 audit of five public benchmarks found that 15.3% of FAIL verdicts were wrong, and those graders were built by careful teams.
What does an RL environment cost?
- Epoch AI reports market prices of $200–2,000 per task, around $300,000 for a Slack-grade replica, and exclusive deals at four to five times the non-exclusive price.
- A cloned website reportedly sells for about $20,000 and is bought once and reused; OpenAI reportedly bought “hundreds” (SemiAnalysis).
- Mechanize argues that tasks cheaper than about $500 waste the compute spent training on them.
In-house, the equivalent costs are engineering salaries, compute for calibration runs and the time of the experts who define what “correct” means.
When does building in-house make sense?
- The workflows or data are proprietary, or you must own the IP outright.
- Security rules keep data or systems from leaving your environment.
- You already own the runtime. Anthropic’s hiring, for example, describes a team that owns the runtime and plugs vendors in.
- Open starting points exist for your domain — Xiaomi reportedly open-sourced 7,780 RL tasks — and open standards such as Harbor and OpenEnv keep your work portable.
When does buying make sense?
- The app category is common — grocery delivery, payments, invoicing, issue tracking — and not a differentiator.
- You need tasks this quarter, not after a hiring cycle.
- You want a verifier that someone other than your own team built and attacked.
- You need expert demonstrations recorded in the environment, and don’t want to run an expert network.
Can you combine both?
Most teams do. Own the runtime and the evaluation harness; buy environments that plug into it; and hold vendors to the same bar as your own team. Insist on standard interfaces, ask for QA evidence per task, and rerun the probes yourself before trusting a vendor’s grades.
Where does Sandworld fit?
Sandworld sells deterministic twins of common app categories with QA you can reproduce: the probes are seeded, so rerunning them gives the same result, and each world’s report is on its page and in the API. Tools come in Claude Messages API, OpenAI function-calling and MCP formats, and every task exports as a Harbor task package (beta) whose verifier stays out of the agent's container, so it can run inside the harness you already own. We’re early-stage — six app twins and one cross-app world, no customers and no certifications yet — so the honest pitch is a small pilot: tell us what you’d build, and we’ll say whether buying beats building for it.