AI NPCs in 2026: How LLM Characters Are Built and Priced

#gamedev #ai #llm #gamedesign

The short answer

An AI NPC is a game character whose dialogue or decisions come from a language model at runtime instead of a fixed script. In 2026 they work best in narrow roles with clear goals: a companion who comments on what you do, a shopkeeper you can haggle with, a suspect you interrogate. They need guardrails, a fallback for when the model is slow or wrong, and a running-cost estimate before launch. Done well, they make a world feel alive. Bolted on, they make it feel like a chatbot wearing a costume.

What has actually shipped

For years AI NPCs were tech demos. In 2025 they started appearing in commercial games. Krafton's life sim inZOI added an experimental "Smart Zoi" setting in which characters' thoughts and actions are driven by a small language model running on the player's own GPU, through NVIDIA's ACE technology; NVIDIA describes it as a 0.5-billion-parameter model. Characters adjust their daily plans based on what happened to them, and players can watch their inner thoughts. NetEase added an on-device ACE teammate to NARAKA: BLADEPOINT MOBILE PC VERSION that fights alongside you, finds items and suggests skills. Other big publishers have shown prototypes, but most AI characters in 2026 are still optional features or experiments rather than the core of a game.

The lesson from what shipped: the successful versions give the model a small, well-defined job inside systems the game already controls.

The five parts of an AI NPC

Input. Typed text, or speech converted to text. Voice feels magical but adds a speech-recognition step, latency and accents to test.
Character brain. A language model given a persona: who the character is, what they know, what they want and how they speak. This instruction block is the most important piece of writing in the feature.
Memory and world state. What the character remembers about the player and what is true in the game right now: quest progress, inventory, time of day. Without it, characters contradict the game.
Guardrails. Rules and filters that keep the character in role, on topic and within your age rating, plus checks on anything the model asks the game to do.
Output and actions. Text or synthesised voice, animation and lip sync, and a short, fixed list of game actions the model may request, such as give item, change mood or open shop. The game validates every request; the model never changes game state directly.

Cloud model or on-device model?

	Cloud model (API)	On-device small model
Response quality	Highest; strong reasoning and writing	Good for short, focused dialogue
Latency	Usually one to a few seconds per reply; streaming helps	Can be faster on a good GPU; slow on weak hardware
Running cost	Pay per token, grows with players	No per-use fee after shipping
Offline play	No	Yes
Build size and hardware	No impact	Adds model weights and memory requirements
Best for	Narrative-heavy PC and mobile games with online play	PC games with capable GPUs, offline play, many NPCs

Many teams prototype with a cloud model, because it is fast to try ideas, then decide per character whether the quality justifies the cost or whether a smaller model will do.

How to estimate running costs

Cloud models charge per token, roughly three-quarters of a word. Each player message sends the persona, the relevant memory and world state, and the recent conversation, then receives a short reply. The estimate is simple arithmetic:

monthly cost ≈ players × conversations per player × turns per conversation × (input tokens × input price + output tokens × output price)

Run it with your own design numbers and your provider's current price list. Three habits keep the bill small: keep the persona tight, summarise long conversations instead of resending them, and use the provider's prompt caching so the unchanging persona is billed at a discount. A cheaper, faster model is often better for in-game chatter anyway, because players notice delay more than eloquence.

Design rules that separate good AI NPCs from gimmicks

Give the character a job."Convince the player not to enter the mine" is a design. "Talk about anything" is a chatbot.
Let the game stay in charge. Quests, rewards and outcomes come from your systems. The model chooses words and requests actions from a short list.
Hide the wait. Thinking animations, short pre-written barks and streamed text make a two-second response feel natural.
Fail gracefully. If the model times out or a filter blocks a reply, fall back to a scripted line in character. Never show an error.
Test like an adversary. Players will try to break the character within minutes. Red-team prompts for jailbreaks, spoilers and off-rating content before launch.

Store rules, ratings and voice

Steam requires developers to disclose AI use in its content survey and to describe the guardrails on live-generated content, and it shows the disclosure on the store page. Age ratings apply to everything a character can say, so your filters must match your rating. If characters speak, check the licence of any synthetic voice, and remember that voice actors' union agreements now include consent and disclosure rules for AI voice replicas.

What it takes to build one

A focused prototype, one character in one scene of an existing Unity or Unreal project, is usually a few weeks of work and the cheapest way to learn whether the idea is fun. A production feature adds memory, guardrails, voice, localisation, cost controls and analytics. We build both on our AI and LLM game development service, using cloud models, including the latest Claude models, or on-device models depending on the game, and quote a fixed price after a short brief.

Sources: NVIDIA: ACE characters in inZOI and NARAKA: BLADEPOINT · VGC on Steam AI disclosure. Checked October 12, 2026.

Originally published on A Plus Games Studio, a game, AR and VR development studio in Sacramento, CA.

DEV Community