Google DeepMind

Genie 2, in December 2024, produced ten seconds of low-resolution world with no real-time interaction. Genie 3, announced in August 2025 and opened to Google US Ultra subscribers through Project Genie in February 2026, sustains a minute of photorealistic environment generated frame by frame as the user moves through it. Google DeepMind states the difficulty plainly: a video model can retroactively adjust frames to keep them coherent, while a world model must stay consistent with the past and with the user's immediate action, without knowing the inputs still to come. The 60-second ceiling is a tradeoff, since world dynamism decays over the sequence and serving costs remain high. The team places itself at 50% of a vision it describes as a simulation indistinguishable from reality. Simmer already uses Genie 3 to train agents that pursue goals in arbitrary 3D worlds.

The same organization ships the plumbing underneath. The Interactions API entered public beta in December 2025 through Google AI Studio, with optional server-side state, background execution for long inference, and native MCP support. Google DeepMind positions it as complementary to generateContent, which stays recommended for standard production responses.

In August 2026, Kevin Hou represented Google as a Core Maintainer of the Agent Plugins 1.0.0 packaging spec, next to Amazon, Cursor, Microsoft, OpenAI and Vercel. Anthropic, whose Agent Skills and MCP servers that format exists to package, is absent from the maintainer list.