Post

A scheduling model built by a 24-year-old in China is assigning work to Claude and GPT. That’s worth more than 4 million views. Multi-model routing isn’t new; what’s new is that someone actually turned it into a product. Just last week, I was manually switching back and forth between several models until I got sick of it—the answer is hidden in the details. I’ll bet Anthropic or OpenAI launches its own official routing layer within a month. Let’s check back this weekend.



More specifically, this scheduling model works like a manager assigning shifts to large language models: instead of asking one model to do everything, it assigns tasks based on task type, cost, latency, and quality requirements. For example, long-form summarization and complex reasoning go to Claude, while code generation, structured outputs, and tool calls go to GPT. When concurrency is high, it prioritizes cheaper models; for critical questions, it steps up to a more capable model. The reason this is worth more than 4 million views is that views show people are interested, while productization shows someone has turned that “interest” into a working workflow. Multi-model routing itself isn’t new—OpenRouter, routers in LangChain, and various agent frameworks can all do similar things—but most remain in the developer glue layer: you have to write your own evaluations, set your own thresholds, and handle failure fallback yourself. The real challenge is the product details: which model performs better on which tasks, how to downgrade after a timeout, whether prompts should be rewritten for each model, how to switch automatically when costs rise by 20%, and whether users can see “why Claude was used instead of GPT this time.” The answer is hidden in these details.

Last week, I wrote a coding request involving a table. I used GPT to generate the structure, switched to Claude to polish the explanation, then switched back to GPT to revise the JSON. I switched back and forth four or five times, until I got sick of it. That’s when I realized users don’t need more models; they need a reliable “dispatcher.” So the real value of this scheduling model built by a 24-year-old in China isn’t showing off—it’s making “scheduling” the default experience: users just describe what they need, and the system handles model selection, result validation, and retries after failures in the background. 4 million views represent attention; productization is what keeps you in the game. I’ll bet Anthropic or OpenAI launches its own official routing layer within a month. After all, model providers know their own costs, latency, and capability limits best, and they have an incentive to bring multi-model collaboration into their official offerings. Let’s check back this weekend.
View Original
post-image
This page contains third-party content and does not constitute any advice, nor does it represent Gate's endorsement of such views. For details, please see disclaimer.

  • 1

Add a comment
Add a comment

Comment 1
ValueValley
Scheduling? I thought it was for flagging crypto shitcoins