𝗜𝗻𝗱𝗶𝗲 𝗴𝗮𝗺𝗲 𝗱𝗲𝘃𝘀: 𝘁𝗵𝗲 𝗲𝘅𝗽𝗲𝗻𝘀𝗶𝘃𝗲 𝗽𝗮𝗿𝘁 𝗼𝗳 𝗔𝗜 𝗡𝗣𝗖𝘀 𝗶𝘀 𝗻𝗼𝘁 𝘁𝗵𝗲 𝗱𝗶𝗮𝗹𝗼𝗴𝘂𝗲, 𝗶𝘁 𝗶𝘀 𝗿𝗲𝘀𝗲𝗻𝗱𝗶𝗻𝗴 𝘆𝗼𝘂𝗿 𝗹𝗼𝗿𝗲 𝗯𝗶𝗯𝗹𝗲 𝗼𝗻 𝗲𝘃𝗲𝗿𝘆 𝗰𝗮𝗹𝗹.

I worked through what an NPC dialogue pipeline would cost on http://B.AI, and one number changed my whole design.

Say your world bible is 50,000 tokens and every NPC line request includes it.

1,000 dialogue calls means 50 million input tokens of the same lore.

On MiMo-V2.6-Flash, at standard prices per 1M tokens:
⤞ Sent as normal input at $0.14, that lore costs about $7.00
⤞ Served as cache reads at $0.0028, the same lore costs about $0.14
⤞ Output stays cheap either way at $0.28 per 1M tokens

𝗦𝗮𝗺𝗲 𝗺𝗼𝗱𝗲𝗹, 𝘀𝗮𝗺𝗲 𝗹𝗼𝗿𝗲, 𝗿𝗼𝘂𝗴𝗵𝗹𝘆 𝗳𝗶𝗳𝘁𝘆 𝘁𝗶𝗺𝗲𝘀 𝗹𝗲𝘀𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗿𝗲𝗽𝗲𝗮𝘁𝗲𝗱 𝗽𝗮𝗿𝘁, 𝗶𝗳 𝘁𝗵𝗲 𝗰𝗮𝗰𝗵𝗲 𝗵𝗶𝘁𝘀.

Hits are not guaranteed, so the FAQ's rule applies: lore and character sheets first, the changing player line last, and no timestamps in the fixed prefix.

The rest of the pipeline I would build on @BAI_AGI:
⤞ Quick barks and ambient chatter on MiMo-V2.6-Flash with deep thinking turned off through thinking.type, since it is on by default
⤞ Quest logic and branching checks on a stronger reasoner, only when a scene needs it
⤞ Gameplay scripting help on GPT-5.4 Mini at $0.75 in and $4.50 out
⤞ Rough trailer and cutscene drafts on Seedance 2.0 Fast, priced per 1M video tokens rather than per clip

One key covers all of it, and the Usage page shows which part of the game is eating the budget.

𝗗𝗲𝘀𝗶𝗴𝗻 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗰𝗮𝗰𝗵𝗲 𝗳𝗶𝗿𝘀𝘁, 𝘁𝗵𝗲𝗻 𝗽𝗶𝗰𝗸 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹.

For a small studio, that order matters more than any benchmark.

The FAQ on caching is here: http://docs.b.ai/llmservice/faq/

@justinsuntron @BAI_AGI #TronEcoStar