There is no single price for AI NPCs, because the three ways of adding them have fundamentally different cost structures, and the structure matters far more than any number a vendor quotes you today. A cloud NPC platform bills you for usage, so your cost scales with how much your players actually talk to your characters. Building your own local stack costs engineering time up front and maintenance forever, but nothing per conversation. On-device middleware sits between the two, trading the build for a licensing fee while keeping inference on the player's machine.
This article explains how each model behaves as your player count grows, which costs teams reliably forget to budget, and how to pick the structure that fits your game. It will not give you competitor price sheets, because usage pricing changes too often for a written number to stay true; check each vendor's current page before you commit.
At a glance
| Cloud NPC platform | Build your own local stack | On-device middleware (Director) | |
|---|---|---|---|
| What you pay for | API usage (tokens, characters, or minutes) | Engineering time, up front and ongoing | A license tied to game revenue |
| When you pay | Every month your game is played | Mostly before launch | After launch, only above a revenue floor |
| Scales with | Playtime per player | Team size and scope | Your game's gross revenue |
| Compute runs on | The vendor's servers | The player's hardware | The player's hardware |
| Works offline | No | Yes | Yes |
| Biggest hidden cost | A hit game multiplies the bill | Retrieval, validation, and upkeep you must build | Player hardware budget |
Prices, free tiers, and hardware requirements move fast; check each vendor's current documentation before budgeting.
What does a cloud NPC platform cost?
Platforms such as Inworld and Convai run the language model on their servers and bill you for what your game consumes, typically metered in tokens, characters, or interaction minutes. The appeal is that you start with almost no infrastructure work, and during development the bill is small because only your team is talking to the NPCs.
The structural risk arrives with success. A usage bill scales with how much your players engage, which means the game you most hope to ship (one where players talk to characters for hours) is the one that costs the most to operate, and the bill continues every month for as long as the servers answer. Free-to-play and live-service games can model this as a cost of goods sold, but a premium game that collects its revenue once and pays for conversations indefinitely has an awkward curve to manage. Cloud platforms also carry an online requirement, which affects certification for offline play and what happens to your NPCs if the vendor changes terms or shuts down. Scope is the other thing to check before comparing prices: what these platforms meter is conversation, so if your design needs quests, plot, or world state that react to the player, you are pricing a character layer when what you need is a director layer. We compare the two architectures in more depth in LoreWeaver vs Inworld and LoreWeaver vs Convai.
What does building your own local stack cost?
Running a small language model on the player's machine removes the per-conversation fee entirely, and in 2026 the models and runtimes are genuinely good enough for shipped games. The cost moves into your engineering budget instead. The model itself is usually free; the work is everything around it, and our guide to running a local LLM for NPCs walks through that stack piece by piece.
Budget for four things that are not the model: a retrieval and memory system so characters remember previous sessions, a validation layer so output cannot break your lore, a structured-output contract so your engine can act on results safely, and ongoing maintenance as models and runtimes evolve. For a team with a systems programmer to spare, this is months of work rather than weeks, and it competes directly with time spent on the game itself. The result is total control with no vendor in the loop, which for some teams justifies the spend on its own.
What does on-device middleware cost?
Middleware hands you the hard parts of the local stack as a product, so the cost becomes a license instead of an engineering project. LoreWeaver Director, currently in beta, is priced on this model: it is free while you develop and free after launch until your game has earned EUR 100,000 gross, after which the fee is 1% of gross revenue above that floor. There are no per-token or per-conversation charges at any point, because inference runs on the player's hardware rather than on a metered server.
The structural consequence is that cost tracks success instead of engagement. A game nobody buys pays nothing, and a game players sink hundreds of hours into pays the same as one they finish in ten, which is the inverse of the cloud curve. The trade-off you accept is the same one as the self-built stack: the model spends the player's compute, so your minimum spec has to reserve memory and GPU headroom for it, and players on very old hardware are outside the audience. The middleware category as a whole is covered in what an on-device AI game master actually is.
Which costs do teams forget?
- The success case. Model the bill at your optimistic player count, not your expected one. Usage pricing that looks trivial at 5,000 players can dominate operating cost at 500,000.
- The player's hardware. On-device approaches shift compute cost to the player. That is free to you in cash terms, but it costs you minimum-spec headroom, and it can cost you buyers at the low end.
- Validation. Whatever generates the text, something must check it against your rules before the game acts on it. Cloud platforms and middleware ship this in different forms; a self-built stack must build it.
- The exit. Ask what happens when you stop paying. A cloud NPC stops talking. A local model you ship keeps working for as long as the game does.
Which should you choose?
- You just want characters players can chat with: that is the easy case, and it is what cloud character platforms are built for; Inworld or Convai will have a talking character running against an API key in an afternoon, and for pure roleplay chat outside a game engine, SillyTavern is free and open source. At that scale the usage bill barely registers.
- You want gameplay that reacts to the player: quests, plot progression, and world state that respond to what the player actually did are beyond a chatbot, however good its dialogue, because conversation is a character layer and this needs a director layer that reads game state and returns validated narrative decisions. That is what Director is built for, in beta today, free until your game has earned its first EUR 100,000.
- You are shipping a live-service game with servers already running: cloud platforms fit your operational model, but compare the bills before assuming they fit your economics, because usage pricing grows with engagement without a ceiling while Director's fee can never exceed 1% of gross revenue above the first EUR 100,000, however much your players talk.
- You have systems engineers and want to own every layer: build the local stack yourself, with open eyes about the retrieval and validation work involved.
- Your problem is actually the lore, not the runtime: structure your world first with Architect, which is free, and decide on a runtime afterwards.
FAQ
Do AI NPCs cost money per conversation? Only on cloud platforms, where the model runs on the vendor's servers and you are billed for usage. Models that run on the player's device, whether self-built or through middleware, have no per-conversation cost.
Is it cheaper to run AI NPCs locally or in the cloud? Before launch, cloud is usually cheaper because free tiers cover development. After launch it inverts: local inference costs nothing per player, while a cloud bill grows with every hour of play for the life of the game.
How much does LoreWeaver Director cost? Director is free during development and after launch until your game has earned EUR 100,000 gross revenue. Above that floor the fee is 1% of gross revenue, with no per-token charges. It is in beta.
What hardware do players need for on-device AI NPCs? Small quantized models run comfortably on mid-range consumer GPUs in 2026, but the model shares memory with your renderer, so the real question is how much VRAM your minimum spec can spare. Requirements shift quickly; test on your actual floor machine.
Do I keep paying after my game launches? With cloud platforms, yes, for as long as players talk to NPCs. With a self-built local stack, no, beyond maintenance. With Director, only the revenue share above the EUR 100,000 floor.
The short version
AI NPC cost is a question of structure before it is a question of numbers: cloud platforms bill you for engagement forever, a self-built local stack bills you in engineering time up front, and middleware converts the build into a license. If on-device economics fit your game, Director runs the whole loop on the player's machine and is free in beta until your game has earned EUR 100,000, and Architect is free today for structuring the world it will run on.