Nobody at a frontier lab has published a price list for RL environments, and nobody will. What exists is a handful of interview-based reports, a few public job listings, two marketplace pages and some funding announcements. This article collects every figure we could verify, with its source, its date, and whether the source saw the number firsthand or repeated it. If a number is not here, we could not source it, and we would be cautious about anyone who quotes it without one.
Per task
The best public source is Epoch AI's FAQ on reinforcement learning environments, published January 12, 2026 from interviews with "18 people across RL environment startups, neolabs, and frontier labs." On per-task prices: "multiple interviewees agreed on a ballpark of $200-$2,000 per task." One environment founder: "I've seen $200 to $2000 mostly. $20k per task would be rare but possible." Epoch adds that the high end "comes up for especially complex software engineering tasks, but it's rare." Firsthand, from the people buying and selling.
Sapphire Ventures' essay on RL environments (April 13, 2026) gives the same shape without an inline citation (its sources list names Epoch AI): tasks "typically ranging from several hundred to several thousand dollars each." Treat it as a second voice, not a second data point.
The compute floor
Mechanize's Cheap RL tasks will waste compute (August 18, 2025) is the one public attempt to derive a price from first principles rather than report one. The arithmetic: a projected 500,000 tokens per transcript (the post puts the 2025 average at about 100,000), a frontier output price of $15 per million tokens, a group of 64 rollouts per task, so "(500,000 tokens × $15 × 64) ÷ 1M = $480 per RL task per run," and "assuming five reuses per task, the total lifetime compute expense will be $2,400 for each RL task." The conclusion: "it's inefficient for AI labs to spend significantly less than ~$500 per RL task," with the authors expecting the floor to rise "roughly fivefold" within a year. The argument is firsthand; the inputs are 2025 prices and a specific rollout count, so recompute it with your own numbers before quoting it.
Per environment
For whole environments the figures are larger and softer. SemiAnalysis, in RL Environments and RL for Science (January 6, 2026, the relevant passage above the paywall), reported that website replicas "often cost about $20,000 per website" and that OpenAI "has purchased hundreds of sites for ChatGPT Agent training and development." Epoch repeats the $20,000 figure citing SemiAnalysis, and adds one interviewee's number for replicas of complex products like Slack: "one interviewee mentioned figures around $300k for these." Sapphire gives the same two tiers in words: "A basic website clone might run tens of thousands of dollars," an enterprise replica "several hundred thousand dollars."
Note what these are: browser and application replicas for computer-use agents. For software engineering environments built from a real codebase, which is what we sell, no public per-environment figure exists. The closest public anchor is the marketplace baseline below.
Contract sizes and exclusivity
Epoch again, firsthand: "Contract sizes are often six to seven figures per quarter." Epoch reports one environment founder noting that contracts are "often seven figures per quarter or more," a researcher at a smaller lab mentioning "contracts in the $300k-$500k range," and two founders independently agreeing that "exclusive deals are roughly 4-5x more expensive than non-exclusive ones." Sapphire: "deals typically land in the 6-7-figure range per quarter."
At the top of the market, TechCrunch reported on September 16, 2025 (updated September 21), citing The Information, that "leaders at Anthropic have discussed spending more than $1 billion on RL environments over the next year." That is a discussed budget, reported secondhand, not a contract. Wing VC's January 2026 analysis estimated Anthropic "spending on the order of tens of millions annually on RL environments" and that OpenAI "signed multiple seven-figure RL-environment and human-data contracts," both the author's estimates with no source given.
What the suppliers earn
Revenue figures for environment companies come from funding announcements and reporting, so they are company-reported or secondhand. Mercor's own blog (May 8, 2026) says "we pay out over $2 million every day to more than 30,000 weekly active contractors." Fortune reported on October 1, 2026 that nine-person Halluminate "crossed the mid-eight figures in annualized revenue run rate" and "counts four top U.S. AI labs as customers." Pulse 2.0 reported on July 9, 2026 that Prime Intellect said demand had scaled to more than $100 million in annualized revenue in under a year. The same TechCrunch piece reported Mechanize "offering software engineers $500,000 salaries to build RL environments." RL List (updated October 5, 2026), which tracks 42 environment companies, states plainly that per-deal pricing figures "are not on the public web."
What the people earn
The contractor platforms publish rates per hour, which is the one place prices are printed rather than reported. Mercor's listing for Experienced Software Engineers building "high performance RL environments" (posted in 2025, now closed) shows "$75-$100 / hr." Handshake's fellow page (updated October 20, 2026) gives hourly rates of "$30-$125." An Outlier listing for coding expertise, closed in July 2025, stated that "pay rates for core project work by coding experts range from USD $25 to $50 per hour." A Terac listing for RL environment engineers on a venture job board (posted August 13, 2026, now closed) showed $120 per hour. These are individual contractor rates in the United States; they are the ceiling a platform pays a person, not what a lab pays a vendor.
Marketplace splits
HUD's guide to selling a codebase (June 24, 2026) is the only published price sheet for software engineering environments we found. Its marketplace, DataVendor, "starts at a $5,000 baseline per codebase and increases with grading," and "The seller receives 80% of the final payout, with 20% covering grading and packaging." Prime Intellect's environments programme (October 27, 2025) pays open bounties "in the $100-500 range" and application-only bounties "ranging from $1000 to $5000+" per environment. Both are firsthand and both are small next to the contract sizes above, which is the point: the marketplace is where an environment proves it grades; the contract is where it earns.
How to read all of this
- Per task, $200 to $2,000 is the only range with multiple firsthand sources behind it. Anything above that needs a reason: a complex multi-service codebase, exclusivity, or a format the buyer cannot produce in house.
- Below about $500 per task the buyer is spending more on rollouts than on the task, by Mechanize's arithmetic. A cheap task is not a bargain for the lab; it is a liability if the grader is weak.
- Exclusivity multiplies price four to five times, by Epoch's interviews. An environment built on a codebase you own is the one you can sell exclusively.
- Marketplace numbers are entry prices, not market prices. A $5,000 baseline and an 80% split get a codebase graded; a six-figure quarter needs a direct relationship.
Our own pricing follows the same logic and is deliberately not on this page: pilots are fixed price, scoped to a defined number of tasks, and graded by the buyer before any larger commitment; per-role rates are quoted on request. The terms are on the RL environments page and the roles on the hire RL engineers page. We will update this article when a new primary source appears; the date on each figure is the date to trust.
Last updated October 26, 2026