A world model gets a meter

Until today, Genie 4 had no price. The October announcement gave none. Yesterday's Project Genie post gave subscribers access but attached its only session-length number, 3 minutes, to the older Genie 3. The developer release attaches the first price and the first session cap to Genie 4 itself, and the numbers are specific enough to plan against.

The billing unit is worth pausing on. Genie 4 is priced in tokens, like a language model, except the tokens count frames: 240 per frame, 24 frames a second. That framing makes a world model look like a utility, and at roughly 10 cents a second the rate is not exotic next to hosted video generation. A full two-minute session costs about $12.10 in video output alone, with input and text output priced separately. An hour of streamed world, summed across back-to-back sessions, would cost about $363 at the same rate. Sessions cannot run that long individually, which is the other half of the story.

Two minutes, and the last stretch is the weak one

The session cap on the developer API is 120 seconds. Inside that cap, Google's documentation adds a warning, not a guarantee: visual stability may degrade past 60 to 90 seconds. The docs give no window of coherence a developer can count on. They also document a shorter-horizon limit that matters more in practice: Genie 4 simulates frames over a rolling temporal window rather than maintaining a persistent 3D scene graph, so looking away from an object for more than about 8 seconds, or turning a full 360 degrees, can cause landmarks, doors or buttons to shift or reset when you come back to them.

The comparison with yesterday's news is awkward for Google, with the surfaces named. In Project Genie, the consumer surface, the older Genie 3 now runs for up to 3 minutes per generation. On the developer API, the newer model stops at two minutes. Google has not explained the difference, and on the published figures the flagship world model gives developers a shorter, metered session than Project Genie gives subscribers.

None of this makes the release minor. A world model behind a standard API key, with a model code and a price, is a different thing from a demo in a research post. Robotics and agent-training teams, one of Google's three named audiences, can now budget real experiments instead of watching launch videos. Prototypers can test whether a steerable world beats a rendered clip for their use case, at about $12.10 a full session rather than a subscription negotiation.

What to check before building on it

The first check is the billing switch. A Genie 4 session bills until the developer calls complete() to end it; Google's documentation warns that closing a browser tab detaches the transport without immediately stopping server-side generation. A forgotten session is a meter still running.

The second check is session economics. A 120-second cap with degradation warnings past 60 to 90 seconds suits prototyping and short interactive scenes. It does not suit long-form world exploration, and stitching sessions together is the developer's problem, not a feature Google announced. Anyone pricing a product on top of Genie 4 should also cost the documented permanence limit: a world whose landmarks can reset after an 8-second look-away is a world whose takes need checking, not just generating.

Sources

  1. [1] Google, "Genie 4 now available through the Gemini API," blog.google, November 19, 2026Read source
  2. [2] Google, Gemini API pricing page (Genie 4 entry: token rates, 240 tokens per frame, approximately $0.10 per second)Read source
  3. [3] Google, Genie 4 documentation (session cap, stability and object-permanence limits, complete() billing instruction)Read source
  4. [4] Google, "Project Genie update featuring Genie 4" (the November 18 consumer rollout and Genie 3's 3-minute sessions), blog.google, November 18, 2026 — as covered in alfw.ai, "Genie 4 Reaches Paying Users at Last," November 18, 2026 (label only, no URL recorded)