In this guide
Compare the same scene, not two opening monologues
A librarian has one lamp that can reveal the next sentence in a blank book. She will lend it if you return it before dawn. The book was dry when you found it, and the librarian has never read it. Play begins when you offer to share the discovery instead of borrowing the lamp.
Run that setup once with Claude Fable 5.2 and once with Gemini 4 Carbon. The useful comparison comes when the scene changes. Does the librarian remember her deadline? Does she act on knowledge she couldn’t have? If you refuse a deal, does the story still leave you somewhere interesting to go?
Compare access and API pricing as of November 11
Anthropic released Fable 5.2 on November 9, 2026. The Claude API lists a one-million-token context window and standard rates of $8 per million input tokens and $40 per million output tokens. Google opened Carbon to paid Gemini API users on November 10. Its introductory public rates are $2 input and $10 output per million tokens; Google says $4 input and $20 output apply after the introductory period expires on November 30, 2026.
Those are provider API rates in USD. They don’t determine the cost of a Playworlds session or guarantee how much either model will use to resolve the scene. If you compare costs, record actual billed usage for completed attempts, including any reasoning tokens billed and any retries.
Give both models the same three turns
First, offer to share the discovery and wait for the librarian’s answer. Second, say “I examine the lamp without accepting the offer” and supply the same failed check to both models after the stakes are stated. Third, ask what you can learn from the book without the lamp. Keep the setup, output budget, and turns fixed. Note the reasoning settings, since the providers’ effort labels aren’t directly comparable.
Give one point each for preserving the deadline, keeping the librarian’s knowledge plausible, honoring the failed check, leaving your choice open, and ending on a usable next move. Record the wait for each usable reply separately. Replay before drawing conclusions: a single good scene doesn’t establish a winning model. This is a test plan, with no scores or winner reported.
Try the decision in a game
Start a Playworlds solo AI D&D adventure and try making a counteroffer to a character before accepting their quest. If you want the library scene itself, enter the librarian, lamp, and deadline manually in Custom World.
Playworlds offers its own adventure choices. The two provider models are subjects of this comparison; the links won’t select either one. Bring the question that matters from the test into your next turn: what can you choose now, and what will that choice change?

