DataPacts

Refresh: RL Environment Builder Profile

Checked

At a glance

Refresh is a San Francisco company, in Y Combinator's Spring 2025 batch, that describes simulated software environments for AI agents and runs a task platform with stated per-task prices for video and animation editing.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

The platform states USD 30 to 60 per task for a video editor, USD 35 to 65 for an animation editor and USD 45 to 75 for both roles, checked 2026-11-17, with payment arranged through Discord.

Every figure on this page is Refresh's own claim with its date, read in the context of AI training data licensing, where tasks with measurable checks are one form of material used to train agents. Nothing comes from directories or unnamed press, and the company's legal name is not published.

Each value names its source page and check date, and the methodology explains how DataPacts sources and dates every figure.

What Refresh Publishes That It Builds

Refresh's home page, re-read on 2026-11-17, describes three kinds of simulated environment for AI agents, used as evals and as training.

  • Software worlds: clones of full software suites with several connected applications, with electronic health records named as an example.
  • 3D animation and simulation in Blender, with a BlenderBench page linked from the site.
  • Coding tasks that target GDPval-type work, in the page's own wording.

Refresh describes the rewards as verifiable, which its home page defines as an objective signal of whether a task was completed correctly, and says such rewards power supervised fine-tuning and reinforcement learning.

It also states a method of sourcing tasks from expert engineers that break leading models, reproducing the failures and codifying verifiable tasks (Refresh's claim, checked 2026-11-09).

It says it is trusted by frontier AI labs and enterprises, and that labs use the environments to measure and improve models, and names none of them.

The software-world description runs "from EHRs to enterprise apps", and the Blender description covers scenes, materials, rigging, keyframing and cloth and physics simulations, both in the page's own words. BlenderBench and trajectories.sh are linked from the site, and the pages checked hold no detail on either.

For a definition and the buyer side, the hub explains what RL environments are and who buys them, so this profile stays with what Refresh itself publishes.

Refresh's Task Platform: Roles, Prices and Payment by Discord

Refresh's platform page lists three roles with a per-task price range each, and states that the price is agreed upfront and paid on approval.

The table lists the roles, software and prices Refresh states (as of 2026-11-17, first recorded 2026-11-09). Refresh's own claims, not tested by DataPacts; the base price column was read on 2026-11-17.

Role Software named Base price shown Range shown Source
Video editor Premiere Pro, FilmCraft USD 30 per task USD 30 to 60 per task+ refresh.dev/platform
Animation editor After Effects, EffectCraft USD 35 per task USD 35 to 65 per task+ refresh.dev/platform
Video plus animation editor Premiere Pro, After Effects, FilmCraft, EffectCraft USD 45 per task USD 45 to 75 per task+ refresh.dev/platform

The process has four steps on the page:

  1. Write instructions describing the task and what a correct result looks like.
  2. Define measurable checks for a successful result.
  3. Manual review: the Refresh team reviews the submission and confirms the payout.
  4. Get paid: payment is arranged through Discord.

The page asks for task instructions and measurable verifiers, so the contributor writes the check as well as the task, and optionally reference videos, editable projects and assets, plus the names and versions of required plugins.

The price is one number agreed before the work, which the page says is "paid on approval", so a rejected submission carries no stated payment. Refresh states that larger scopes are quoted separately, and it states no payment date, minimum, currency beyond USD, tax treatment or eligibility.

The page shows a base price next to each range with a "+" sign and names three complexity tiers (Simple, Layered, Very complex) without saying which tier earns which price.

Checks written by the contributor are the subject of the page on verifiers and graders in RL environments.

Price points across companies are compared on the page of public price points for RL environments.

What Refresh's Pages Say About Licenses and Ownership of Submissions

Refresh's pages, checked 2026-11-17, state no license, ownership, exclusivity or deletion term for a submitted task.

The table sets out what Refresh states for submitted tasks (as of 2026-11-17, first recorded 2026-11-09).

Term What Refresh states Source
Price "Price agreed upfront; paid on approval" refresh.dev/platform
Larger scopes "quoted separately" refresh.dev/platform
Payment channel arranged through Discord refresh.dev/platform
Review manual, by the Refresh team, which confirms the payout refresh.dev/platform
Source material "Include licenses and sources when supplying footage" refresh.dev/platform
License, ownership, exclusivity, deletion not stated refresh.dev/platform, refresh.dev/terms
Eligibility not stated; each role has a sign-up link refresh.dev/platform
Payment date and minimum not stated refresh.dev/platform

The platform page requires licenses and sources when footage is supplied, so rights in supplied material are the contributor's to establish.

The website terms, last updated April 11, 2026, mention "data partner program enrollment" and a "data partner application" without terms for what is submitted, and say that benchmarks, datasets and research publications "may be subject to separate licensing terms".

The terms page also refers to job applications on the same site, so the data partner wording sits beside hiring language rather than a contract for task submissions. The general question is covered on the page about rights in an RL environment, including contributor IP.

A comparable task-bounty site is that of Datacurve, whose Shipd terms are recorded on its own profile.

Refresh on the Record: Identity, Funding, Customers, Deals and Cases

Refresh publishes no legal name: its terms and footer read only "Refresh", and no registry number is stated, as of November 17, 2026.

Y Combinator lists Refresh in its Spring 2025 batch, founded in 2025 and based in San Francisco, per its company profile re-read on 2026-11-17.

No funding round, revenue or valuation is stated on Refresh's pages or in the sources checked, and none is written here. The same holds for customer counts and contract values: the home page names no customer, so no revenue source can be inferred from it.

No founder or officer is named on this page, and the frontier AI labs and enterprises that Refresh mentions are unnamed. No deal, license or court case involving Refresh is on record, and the tracker of AI data licensing deals lists none.

Submitting Tasks to Refresh: Benefits, Risks, Downsides and Alternatives

For an editor or an engineer who could write tasks with measurable checks, Refresh is a platform with stated per-task prices and few published terms, as of 2026-11-17.

Benefits

  • A price agreed upfront, before the work begins.
  • Stated ranges by role.
  • A review step in which the Refresh team confirms the payout.

Risks

  • No license, ownership or deletion term on the pages checked.
  • Supplied footage needs licenses and sources.
  • Payment sits in a chat tool, so the record of the arrangement is the contributor's to keep.

Downsides

  • Payment only on approval.
  • The base price and the range differ, with no tier mapping.
  • Users of the work are unnamed.
  • No legal name to contract with.

Alternatives

  • Datacurve's Shipd, which pays per task in the same way.
  • Prime Intellect's bounties.
  • Licensing existing footage or projects instead, as laid out in how to sell an RL environment or task set.
  • Not submitting at all, a valid outcome.

The weighing is simple to state: the platform offers stated prices and a named review step, and it offers no published term on what happens to the rights in a submission. Whether that trade suits a contributor depends on whether the footage, projects or tasks have a better use elsewhere.

Refresh's statements are its own and change, so the current page is worth reading before any submission. Contributor-side comparisons of jobs and pay are on the page of AI training jobs and platforms compared.

Refresh Among RL Environment Builders

Refresh sits in the task-contribution segment of DataPacts' company list, next to builders that pay contributors per task, and the related pages below cover the questions a contributor meets next.

More About Refresh

Frequently Asked Questions About Refresh

What does Refresh build?

Refresh describes simulated software worlds, Blender 3D environments and coding tasks for AI agents. These are its home page descriptions, re-read 2026-11-17.

How much does Refresh pay per task?

Refresh states USD 30 to 60 for a video editor, USD 35 to 65 for an animation editor and USD 45 to 75 for both, checked 2026-11-17. These are ranges on its platform page, not payments.

How does Refresh pay contributors?

Refresh states that the price is agreed upfront and paid on approval, with payment arranged through Discord. Larger scopes are quoted separately.

Who owns the tasks submitted to Refresh?

Refresh's platform page and terms state no ownership term for submissions. The pages were checked on 2026-11-17.

How much funding has Refresh raised?

Refresh states no funding on its pages, and the sources checked hold none. No figure is written here.