How AI NPC Platforms Charge - Pricing Models Compared

AI NPC platforms charge in one of five ways: per response, per minute, per token, by making the player pay, or through a monthly plan with included credits. The sticker price tells you very little. What matters is what is being counted, because that decides what happens to your bill when players actually start talking.

This guide walks through each model, runs one small game through all five, and lists the costs that never show up on a pricing page.

Costs at a glance

How each AI NPC pricing model behaves when players show up

Per response

Convai counts an interaction every time a player sends input and the character answers, in text or voice. Each plan has a monthly interaction quota, with a smaller cap on replies from flagship language models (Convai's own example is 3,000 interactions a month on Indie Dev, 1,500 of them on flagship models) and a separate cap on high-quality third-party voices, which stop being available once it is reached.

Two limits matter more than the quota. Character concurrency is set per API key: on a limit of 1, one player at a time can talk to your NPCs on that key, however many NPCs you have. Larger plans also cap monthly active end users. Quotas reset each billing cycle, and unused interactions do not carry over.

Per minute

ElevenLabs bills agent conversations by the minute. Each subscription includes a block of minutes, extra minutes cost more, and the language model's usage is passed through separately. Calls made while setting up and testing an agent are billed at half price, so development is not free. Long silences in voice calls are discounted.

Charisma.ai sells pay-as-you-go credits, priced so that one pack covers roughly 200 experience minutes. Enterprise projects move to a fixed monthly fee based on estimated usage, plus a development fee.

Per-minute pricing suits short voice interactions. It gets expensive in games where players hang around and chat.

Per token or per character

OpenAI prices its realtime voice models by audio and text tokens, with separate rates for input, cached input and output. There is no per-minute rate for a conversation, so you have to measure your own sessions to know what one costs.

Inworld bills against a dollar credit balance: text-to-speech per million characters, speech-to-text per hour of audio, and language models per million tokens. Plan credits roll over for a limited time. This is the most granular model and the hardest to forecast before launch.

Player pays

Player2 gives players free models and daily credits through their own Player2 account, and its SDKs sign each player in through the Player2 app or a browser. You pay nothing for inference, but every player meets a third-party sign-in before your characters can talk, and what happens when their allowance runs out is between them and Player2.

Monthly plan with credits

You pick a plan, it includes credits, and you upgrade when you need more. You know your maximum spend before launch and choose when it changes.

One small game, five bills

Take an indie game with 500 players a month. Each plays three sessions and has about 15 exchanges with characters per session. That is 22,500 replies a month.

• Per response: 22,500 interactions against your quota. The bigger question is how many of those 500 play at the same moment, because concurrency is capped per key.
• Per minute: if each exchange takes about 20 seconds of speech in both directions, that is roughly 7,500 conversation minutes, plus whatever the language model costs on top.
• Per token: every one of those 7,500 minutes becomes input and output audio tokens. A chatty player costs several times a quiet one, and you only learn the average after launch.
• Player pays: nothing to you, and 500 players asked to create a Player2 account first.
• Monthly plan: a known monthly ceiling. If the game grows, you move up a plan on your schedule.

The numbers are made up. The shape is not: usage-based models make your margin depend on how much your players enjoy talking, which is the one thing you are trying to increase.

Costs that never appear on the pricing page

• Testing. Writers and designers talk to characters hundreds of times before launch. Check whether that is billed.
• A key server. Raw voice APIs expect your own backend to issue short-lived keys so a secret never ships in the client. That is a server you build, host and pay for.
• Speech-to-text. Some vendors bill transcription separately from replies.
• Voice tiers. Premium voices can sit behind separate caps or rates.
• Concurrency upgrades. A launch spike can push you into a higher plan for a single weekend.
• The character layer. On a raw model API, memory, personality, guardrails and game actions are engineering time you pay for in salaries, not invoices.

Questions to ask any vendor

1. What exactly counts as one unit: a reply, a minute, a token, a player?
2. Are there caps on concurrent users or monthly active users?
3. Does unused usage roll over?
4. Is testing during development billed?
5. What happens when a quota runs out mid-session?
6. Can you run locally or self-host if costs get out of hand?

How Neocortex charges

• Monthly credits are included in every plan, and commercial use starts on Pro.
• Free testing: talking to your characters in the browser editor does not use credits, on any plan.
• Usage gating tools in the Unity SDK, so you can meter how much each player uses and decide what happens at the limit.
• Local prototyping: you can try text chat with open models running locally through Ollama in Unity, which costs nothing per message.

Start with the free plan

The cheapest way to understand any pricing model is to build one character and watch what it consumes. Build yours on the free plan, test it in the browser as often as you like, and move to Pro when you are ready to ship a commercial game.

Pricing details were last checked in October 2026 and change often. Always confirm on each vendor's pricing page.

Frequently Asked Questions