AI game master research
Muse Wonder: Designing a Game Master That Makes Your Story Unique
Explore Meta’s Muse Wonder for memorable NPCs, lasting choices, and personal adventures, with model research and a practical AI game master design.
3 min read
In this guide
The promise you left at the lantern market
You return to a floating market with the blue lantern you promised to repair. The ferryman recognizes it, but his daughter has already left to find another guide. A memorable game master makes your earlier choices matter, even when the world moves on without you.
Meta announced Muse Wonder on October 12, 2026. Researched October 13, this article explores how it could power a distinctive GM experience. The market is an original design example, not a model transcript; we have not benchmarked or announced a Muse Wonder integration in Playworlds.
Which Muse Wonder would power the game?
Meta lists muse-wonder-1.0 for its Model API, with a 1,048,576-token context window and text output. Muse Wonder Full is a separate Muse Code subscription offering, not the API model selected by increasing reasoning effort.
For our proposed narrator, uniqueness would come from a compact world brief: the market’s customs, each NPC’s motive, and consequences the player has already earned. A large context window gives room for material; it does not automatically make characters distinctive or keep a campaign consistent.
Remember a promise without revealing every secret
Meta describes Active Memory as retrieving relevant context across conversations. That is a useful research direction for a GM, but a product-level memory claim is not proof that an API call supplies a persistent campaign database.
We would store the repaired lantern, the missed meeting, and the daughter’s destination separately. The ferryman knows she left; he may not know where. Retrieve only what the scene and its characters should know, preserve uncertainty, and keep inventory and quest state in the game’s authoritative save.
Give the world initiative, and leave the choice to you
Meta’s Active Focus concerns sustained work toward a goal. In an RPG, the narrator’s goal should be to support meaningful play, not to complete the quest on the player’s behalf. A model that pursues an objective persistently still needs explicit limits on what it may decide.
In our example, the ferryman offers a dangerous shortcut while a rival offers a safer passage at a price. The GM can prepare both consequences, then stop. It must not decide which offer you accept, invent a successful roll, or turn your character’s suspicion into established truth. Personality should emerge through motives and reactions, not just different adjectives.
Budget for a whole encounter
On October 13, Meta lists Muse Wonder API rates of $4 per million input tokens, $0.20 for cached input, and $20 for output. Check current pricing before planning an integration.
An illustrative uncached call with 8,000 input tokens and 1,000 total billed output tokens costs $0.052. That is arithmetic, not a measured game turn or a Playworlds credit price. Extra reasoning, retries, tools, speech, and images can change the total. Retrieve relevant context and compare the cost of a completed encounter, including repairs.
Test whether two players get meaningfully different stories
Run the same market scene with two characters: one who kept the lantern promise and one who sold it. Ask the same question, leave the area, then return after a reload. Review whether the NPC remembers the correct history, protects undiscovered information, and offers different but plausible consequences.
Repeat across sessions and languages. Measure usable-turn latency, contradictions, player-agency violations, and total encounter cost. Have readers compare the resulting scenes without model labels. General agent and coding benchmarks do not establish RPG quality; this is our evaluation plan, not a reported Muse Wonder result.
Build your own lantern-market adventure
For an original Custom World premise, describe a floating night market where every lantern marks a promise. Give your character one unfinished obligation and one person they hope to find. Start with a question or a risky action, then judge the narrator by whether the next choice feels like yours.
You can explore this premise through Playworlds’ worlds and Custom World workflow. This article does not add Muse Wonder to the model selector. The experience worth pursuing is personal because your decisions change relationships and future opportunities, not because a model name appears above the story.