Daytona at $1B, Xiaomi open-sources 7,780 RL environments
Daytona is selling disposable agent sandboxes at a $1 billion valuation, and Xiaomi has open-sourced 7,780 RL environments. The environment business is not collapsing; it is re-pricing in layers. Generic content is being commoditized, while exclusivity, runtime scale and acceptance testing are going up in price.
01The price list for one RL task
Epoch AI's estimates from earlier this year are the most complete public reference. A single RL task usually costs $200 to $2,000, and exceptionally complex software engineering tasks reach $20,000 per task. Environment contracts are signed quarterly, at hundreds of thousands to millions of dollars per deal. One less-noticed number: running RL training on one task costs about $2,400 in compute. The compute to train on a task costs more than buying the task itself.
A second number shows the market structure. In early October, a public registry tracking data transactions recorded 159 suppliers bidding on 38 procurement projects from 29 companies. micro1 sells enterprise data partnerships in three tiers — $100,000, $500,000 and $1 million. Handshake's projects span $100,000 to over $4 million. To attract supply, micro1's bounty for each signed enterprise profile rose from $25,000 in June to $50,000 in October. A doubling bounty means the bottleneck is on the supply side, not the demand side.
02Sandboxes became a standalone business
Daytona in one sentence: a disposable virtual computer for every agent, started in 60 milliseconds. Two customer numbers are worth remembering. Snorkel ran over 1 million sandboxes on it and cut its evaluation cycle by 96%. Trajectory ran 700,000 and pushed environment startup time from 30 minutes down to seconds. The company's own disclosed revenue figure: annualized revenue from January to September grew 25x, with the base not disclosed, so treat the multiple as a multiple.
Why can a sandbox be split out of the environment and sold separately? The answer sits in the engineering requirements of RL training. A Chinese-language engineering retrospective on repo-level agent training, circulating this week, puts it plainly: environments must be resettable, isolated and parallelizable — a single rollout corrupts repository state, and resets must happen in seconds or training throughput collapses. If tests have randomness, the same code passes sometimes and fails others, and noisy rewards destroy training. With hundreds of containers running in parallel, resource contention produces false negatives. The author's conclusion: engineering complexity far exceeds algorithm complexity. That is exactly why Daytona's customers pay. Reset speed and isolation quality are not ops details; they directly determine the quality of the trained model.
03Which price open source actually cuts
Now Xiaomi's 7,780 environments. They cover five domains: programming, cybersecurity, knowledge work, web development and music. One example shows the quality bar: one environment fully reproduces the reinsurance financial due-diligence process of Argo Group. The agent plays an analyst and reconciles numbers across a simulated SAP reinsurance system, a receivables CSV and a financial model, then explains why the figures do not match. This is not a LeetCode-style question bank; it is a workflow with business systems attached.
Will open source kill the environment business? The founders building these environments say no. Savante AI founder Sachin Dharashivkar told AIM that the repository's value is showing people how frontier teams construct environments, but every model company needs different environments — if everyone trains on the same batch, the resulting capabilities have no differential. Epoch's data supports this: exclusive environments cost four to five times as much as non-exclusive ones. Open source has cut the price of only the "generic, reusable environment" segment, which was never where the profit was.
Prime Intellect is the other reference. Its Environments Hub is itself open source; the company charges for hosted training, evaluation, and secure sandboxes that run thousands of environments concurrently. Content free, runtime and acceptance paid — the same split Daytona makes from the layer below.
04How buyers should run the numbers
Put the three layers together and the buyer's math is clear. Full cost per task has three parts: content at $200 to $2,000, with complex tasks priced separately; about $2,400 of training compute per task; plus environment operations and acceptance. Compute is the biggest item, so sandbox reset speed and concurrency decide the total bill. Contract that part on performance metrics, not on seats or duration.
On content, the right use of open-source repositories is calibrating quality expectations, not training. Retraining on environments others have already trained on buys no differentiation. Use them to build a reference for what a qualified reinsurance due-diligence environment looks like. What is worth paying for is exclusive tasks and their verifiers, with exclusivity written into contract terms — per Epoch's framing, the four-to-five-times premium is the market price.
Acceptance remains unavoidable. Once environments are cheap, defects in batch delivery are also batch-sized. The hack rate and verifier misjudgment rate discussed last week have to be re-run at every layer, including the 7,780 open-source environments. Open source lowers the cost of trial and error; it does not lower the cost of acceptance.