How we build and bill for AI, what makes our models different, and why the standard approach falls short.
Our Game AI Platform is delivered as a SaaS product with API access. You connect to it, configure it for your project, and pay a subscription based on usage. The platform handles code generation, graphics generation, playtesting, and engine integration without requiring any custom model work on your end.
For clients who need something the platform doesn't cover, we take on custom engineering projects. These engagements involve building and adapting models to your specific use case, and they're billed differently. The rest of this page explains how that billing works and why our approach to custom models is different from what most providers offer.
When you use a mainstream AI subscription, you pay around $200 per user per month. That sounds straightforward, but what you're actually buying is an invisible quota of tokens. Hit the quota and you face multi-hour lockouts, throttled responses, or a quiet downgrade to a lesser model.
The alternative is the token model, where developers connect directly to the API and pay per token consumed. There are no caps, but also no guardrails. A poorly optimized agent loop running over a weekend can rack up thousands of dollars without producing anything usable. Both models put the risk and the cost of inefficiency entirely on you.
Input tokens, what the model reads, are comparatively cheap. Output tokens, what the model writes, cost three to five times more because they must be generated one at a time. When you move from simple chat to agentic workflows, where an AI autonomously loops through tasks and re-reads its own history at every step, token consumption multiplies rapidly. Neither billing model gives you a way to budget predictably for that.
We charge for the workflow, not the tokens behind it. When we build you a custom application, the cost of running it is ours to manage. We optimize the prompts, structure the agent loops, and eliminate wasteful patterns behind the scenes. You pay a predictable price for the output.
This aligns our incentives with yours. When we make the system more efficient, our margins improve and your costs stay flat. When your usage grows, you scale within a pricing structure you can budget for, rather than watching a meter spin unpredictably.
The reason this works is that we control the full stack. We build the model, we optimize how it runs, and we host it on hardware we operate. The efficiency gains from that control are what make predictable pricing possible.
Most AI providers give every customer the same underlying model and differentiate only at the application layer. The model itself never changes. No amount of prompt engineering can teach a model new domain knowledge or change its fundamental behavior.
We work below that layer. We adapt the models themselves to fit your business through fine-tuning, continued pretraining, retrieval-augmented generation, and structured prompt engineering with DSPy. Because we shape the model to your use case, the system is more efficient by design. It doesn't need to be told what your industry terminology means every time it runs. It already knows.
That efficiency is what makes workflow-based pricing viable, and it's why a customized model will consistently outperform a general-purpose one with a carefully written prompt.
We are not giving you access to someone else's model with a different label on the interface. We are building and adapting models for your specific use case, running them on infrastructure we control, and billing you for the outcome rather than the token count.
