OpenReward — 330+ RL Environments through one API (General Reasoning announcement)

TL;DR

OpenReward (General Reasoning, launched 24 Mar 2026) is a managed platform serving 330+ RL environments as API endpoints, backed by 4.5M+ unique RL tasks and autoscaled sandbox compute. Each environment exposes a standard interface — the Open Reward Standard (ORS), an extension of Anthropic’s MCP with RL primitives (episodes, reward signals, task splits, curriculum management) — so environments published once can be consumed by any trainer (Tinker, Miles, Slime are named). It is the most aggressive “environment-as-managed-service” datapoint to date and is the natural endpoint of the same trajectory occupied by SETA: Scaling Environments for Terminal Agents (env repo) (terminal, ~400 tasks, single-host harness) and Toolathlon-GYM: Large-Scale Long-Horizon Environments for Tool-Use Agents (MCP enterprise, 503 tasks, local Docker Compose). Tweet-as-primary because the artifact is a product launch with no underlying paper.

Key claims

330+ RL environments served behind one HTTP API, with task-level granularity (4.5M+ unique tasks across them) and autoscaled sandbox compute provisioned per rollout [tweet body].
Underlying protocol is the Open Reward Standard (ORS), framed as “MCP for RL” — episodes, reward signals, task splits, curriculum management as first-class primitives [docs.openreward.ai].
Trainer-agnostic by design: explicitly compatible with Tinker, Miles, and Slime; environments are versioned packages published once and consumed by any trainer [tweet body, PyPI 0.1.81 README].
Per-environment compute is provisioned as sandboxes (e.g. 0.5 CPU / 1 GB RAM, Python 3.12 data-science image, network access optional) with the platform handling scaling [openreward.ai/GeneralReasoning/AirlineRM, openreward.ai/GeneralReasoning/Skywork-OR1-RL-Data].
Existing environments include Skywork-OR1-RL-Data (~119K math + code problems with math_verify and sandboxed code execution as verifiers) and synthetic long-horizon decision environments like AirlineRM (30+ multi-day fare/overbooking/disruption-management turns with dense per-day rewards) [openreward.ai environment pages].
Environment authors write a Python Environment subclass exposing tools as @tool-decorated async methods on Pydantic input models; a Server wraps it in a FastAPI app exposing the ORS over HTTP with SSE streaming; agent rollouts are logged back to OpenReward for analysis and training [PyPI 0.1.81 README].
Environments are deployed via the standard git→build→serve loop: push the env package to GitHub, connect the repo in the OpenReward dashboard, configure CPU/memory/scaling, every push triggers an automatic build and deployment to the username/environment-name namespace [PyPI 0.1.81 README].

Method

OpenReward sits one level above the per-task environment repos that have been filed in the last week. Where SETA (SETA: Scaling Environments for Terminal Agents (env repo)) is a 400-task Terminal-Bench-compatible dataset that you Docker-compose locally, and Toolathlon-GYM (Toolathlon-GYM: Large-Scale Long-Horizon Environments for Tool-Use Agents) is a 503-task MCP/PostgreSQL stack you Docker-compose locally, OpenReward removes the “you Docker-compose locally” step: environments live in OpenReward’s autoscaled fleet, identified by org/env-name, and the trainer talks to them over HTTP via the ORS protocol. The protocol itself extends MCP — MCP standardizes tool/resource invocation for a chat model; ORS adds the loop primitives an RL trainer needs (episode lifecycle, reward emission, task splits for train/val, curriculum sampling).

The platform-side artifact has three pieces: (a) the ORS specification at openrewardstandard.io; (b) the openreward Python SDK on PyPI (env-authoring + trainer-side rollout client) shipping orwd init scaffolding for a Docker-sandboxed env template; (c) the hosted dashboard at openreward.ai where environments register their git repos and configure resource limits. Each rollout returns a structured trajectory with per-step tool calls, rewards, and termination state, which the trainer (Tinker/Miles/Slime/other) consumes.

Results

No model or benchmark numbers — this is a platform launch. The pertinent quantitative claims are operational: 330+ environments live at launch (vs. SETA’s 400-task single-environment, Toolathlon-GYM’s 503-task single-environment), 4.5M+ aggregate unique tasks across them, ORS-compliant trainer integrations announced for three named RL frameworks (Tinker, Miles, Slime).

Why it’s interesting

This is the “PyPI / HuggingFace moment” for RL environments. The pattern across the last week of filings — SETA: Scaling Environments for Terminal Agents (env repo), Toolathlon-GYM: Large-Scale Long-Horizon Environments for Tool-Use Agents, VisGym: Diverse, Customizable, Scalable Environments for Multimodal Agents, vlm-gym: RL gym for vision language models (JAX) — is environment-as-package: a per-task verifiable Docker image with task.yaml / evaluation.py / sandbox. OpenReward takes the next step and turns environment-as-package into environment-as-managed-service, with a shared protocol decoupling env author from trainer. For Luma’s research team, this matters in two ways. First, the SETA “Remote Environment Manager” idea (run rollout containers on a separate fleet because 256-concurrent-Docker on the training host fails) is now a productized assumption rather than per-project plumbing — which complements MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling‘s streaming-rollout queue at the trainer side. Second, the ORS-extends-MCP framing is the same “shared interface across heterogeneous backends” recipe that lets Tool-Use Agents post-training pipelines drop a new env in without re-plumbing reward gathering — and if it works at this scale it cuts a meaningful fraction out of the engineering cost of every future “GRPO on a new domain” project.

A reasonable conflict signal vs prior filings: SETA argues for frontier-model-synthesized tasks scoped to a single benchmark (Terminal-Bench), and Toolathlon-GYM argues for brand-obfuscated, ground-truth-anchored tasks to prevent shortcut behavior. OpenReward’s curated mix (Skywork-OR1 math/code on one end, synthetic AirlineRM-style long-horizon decisions on the other) is more catalog than benchmark; whether this catalog model actually trains good models depends on (a) how task quality is policed across 330+ third-party-authored environments and (b) whether the ORS abstractions leak enough that trainers still have to special-case each env. Neither is addressed in the launch tweet.