OpenAI Gym's foster home grew up and got a PhD

What it does

Gymnasium is the continuation of OpenAI’s Gym: a Python API that lets reinforcement learning algorithms and environments talk to each other through a simple env.reset() / env.step() contract. It ships with reference environments spanning toy text problems, classic control, Box2D physics, MuJoCo robotics, and Atari 2600 emulation. Third-party environments plug into the same API.

The interesting bit

The Farama Foundation took over maintenance after OpenAI moved on, and rather than letting it rot, they formalized the interface with an arXiv paper and strict environment versioning (CartPole-v1, not just CartPole). That versioning is the boring part that keeps RL research reproducible — when an environment changes in ways that could alter learning curves, the version number ticks up.

Key highlights

• Standard env class interface: reset(), step(), close() with terminated/truncated flags
• Bundled environments from trivial debug toys (Toy Text) to MuJoCo multi-joint physics
• Atari support via ALE emulator; third-party ecosystem with compatibility shims
• Python 3.10–3.14 tested on Linux and macOS; Windows PRs accepted but not officially supported
• Companion ecosystem includes PettingZoo (multi-agent) and CleanRL (reference implementations)

Caveats

• No official Windows support
• Optional environment dependencies (Atari, MuJoCo, etc.) are split out and can be finicky to install; gymnasium[all] pulls in “a massive number”

Verdict

If you’re doing single-agent RL in Python, you’re probably already using this or something built on it. If you’re not in RL, it’s a well-designed API contract worth studying — and otherwise not your concern.