For low-cost NPC dialogue in a browser game, DeepSeek and Qwen are the cheapest per-token API options. Call them from a backend, not directly from the client, and cap context length so token cost per exchange stays small.
At a glance
| Fact | Value | Source |
|---|---|---|
| Cheapest NPC dialogue APIs by input token price | DeepSeek, Qwen | intelli-verse-x.ai |
| DeepSeek context window | 128K tokens | intelli-verse-x.ai |
| In-browser inference engine using WebGPU | WebLLM | github.com |
For NPC dialogue in a browser game, DeepSeek’s API and Qwen are the cheapest per-token options among hosted LLMs, both priced well below mid-tier models like Claude 3.5 Haiku or GPT-4o Mini (source). DeepSeek offers a 128K context window; Qwen’s rate limits vary by tier, and budget APIs generally cap requests per second lower than premium tiers.
If you build your browser game with an AI coding agent, the remote Playgama MCP server connects tools like Cursor or Claude Code directly to your developer cabinet to manage forms, builds, and sandboxes.
The browser-specific decision is where the call happens. Never put an API key in client-side JavaScript that ships in your build – anyone can read it from devtools or the page source. Route dialogue requests through a small backend (a serverless function works fine) that holds the key and forwards prompts; the same pattern AWS recommends for game backends generally – keep credentials off the client (source).
If you want zero server cost and are willing to accept a large download, WebLLM runs models locally in the browser via WebGPU, with no server calls after the model is cached (source). This avoids per-token cost but adds a first-load download and requires a WebGPU-capable browser, which some mobile and portal iframe environments won’t reliably have.
Watch context growth: sending full conversation history to keep an NPC’s memory coherent multiplies token cost with each exchange, so trim or summarize older turns rather than resending everything.
Sources
- Cheapest LLM API 2026 | Game AI & NPC Pricing Guide
- How hosting with Amazon GameLift Servers works
- GitHub – mlc-ai/web-llm
- LLM-Driven NPCs: Cross-Platform Dialogue System for Games
- Playgama Bridge SDK: getting started
Related questions
Can I run an LLM fully in the browser with no server calls?
Yes, WebLLM runs models locally using WebGPU with no server support after the model is downloaded and cached, but it needs a WebGPU-capable browser and adds a first-load download.
How do I stop NPC dialogue costs growing over a long conversation?
Sending the entire conversation history each time keeps memory coherent but increases tokens per call; summarize or trim older turns instead of resending everything.
Is it safe to call the LLM API directly from client-side game code?
No, an API key embedded in browser JavaScript is readable by anyone; route requests through a backend that holds the key, similar to how game backends generally keep credentials off the client.
Last updated: 30 September 2026