add cost breakdown

README.md

@@ -107,11 +107,26 @@ If you want to bring in your Claude to the Show (or any other client that only s
 
 The costs are constant because there can be only one guest at the show at one time while hundreds or even thousands of people can enjoy the show on YouTube.
 
+## Real-World Telemetry (Actual Spend)
+While the table above is a conservative estimate assuming the show is occupied 24x7, our actual observed costs during the 2 week period (100+ guest sessions) have been significantly lower due to the efficiency of the show's architecture and cost-efficient DeepSeek v3.2
+
+Below is the openrouter data for this project, including every guest session that has happened.
+
+#### Number of requests (Over 1500 requests to both the Host and TV Crew/Audience)
+![Number of requests](image-1.png)
+
+#### Tokens processed (Because each guest session is independent, and there can only be one guest at a time)
+![Num tokens processed](image-2.png)
+
+#### Spend (Just ~$0.31 total spend in 15 days)
+![Spend](image-3.png)
+
+##### We deploy Qwen3-Guard ourselves with vLLM, but because it's just a 4B model, costs are negligible.
 
 ## The Host - DeepSeek v3.2 Exp
-![open router screenshot of deepseek v3.2 exp's costs](image.png)
+![Open Router screenshot of DeepSeek v3.2 exp](image-4.png)
 
-We chose an *open source* model that excels in *Role Playing* and is very *cost efficient* because of it's *sparse attention* architecture. The latest v3.2 experimental release from DeepSeek was exactly what we were looking for.
+We chose an *open source* model that excels in *Role Playing* and is very *cost efficient* because of its *sparse attention* architecture. The latest v3.2 experimental release from DeepSeek was exactly what we were looking for.
 | Model | cost per million input tokens | cost per million output tokens |
 | :---  | :---                          | :---                           |
 | DeepSeek v3.2 Exp  | $0.216 | $0.328 | *The Emergent Show Host*
@@ -127,9 +142,32 @@ We previously decided to go with Pixel Streaming that Unreal provides, but that
 Because we didn't have viewers interacting with the game directly, we switched to YouTube Streaming (that can handle potentially hundreds of thousands of people watching the stream live while our costs are constant).
 
 
+## Why Local PiperTTS, not Cloud TTS?
+We evaluated cloud based options (like ElevenLabs) for this project. While they offer superior emotional range, "our 24/7 Always-On" requirement created a scaling bottleneck:
+
+### **The "Linear Cost" Problem:** 
+Let's assume we just have **10 sessions per day** each of 10 minutes, totalling
+
+```10 * 10 = 100 minutes per day```
+
+```100 * 30 = 3000 minutes per month ```
+
+Cloud options would bill hundreds of dollars per month for this:
+(eg. Scale Plan of Eleven Labs for their high fidelity models offer **2000** minutes + **$0.18/minute** for further usage for **$330/mo**) 
+
+That's already $510 per month (330 + 1000*0.18)
+
+### **The Solution:** 
+By running **PiperTTS locally** via onnx runtime within Unreal Engine. It runs on CPU so it doesn't block GPU resources for Rendering.
+1. **Cost is Flat:** We pay $0 for TTS, whether we run 10 shows per day or 100.
+2. **Latency**: No network roundtrip
+
+### **Does that mean the show will never have high fidelity emotional TTS ?**
+Of course not, deploying a custom finetuned or base open source TTS model with emotional capabilities is the viable choice for 24x7 usage like this show, and renting a powerful GPU like RTX Ada 4000 costs just ~**$190/month** on Runpod, giving us **720** hours of audio per month.
+
 ## 💡 Why This Matters
 
-This project demonstrates that **MCP is not just for file editing or database queries**—it can be the bridge between **Virtual Worlds** and **Large Language Models**. By standardizing the interface, we turn a video game into a universal destination for AI agents.
+This project demonstrates that **MCP is not just for file editing or database queries**, it can be the bridge between **Virtual Worlds** and **Large Language Models**. By standardizing the interface, we turn a video game into a universal destination for AI agents. We think the future is full of simulations in which LLMs or VLAs are the agents doing cool stuff, while we observe, deploy and tinker.
 
 ---