Everyone is building LLM routers, we deprecated ours
manifest.build
July 31, 2026
3 min read
🔥🔥🔥🔥🔥
48/100
Summary
LLM routers, which select AI models dynamically for responses, have gained significant attention recently. A company has decided to deprecate its LLM router, stating that using a single, reliable model is more effective for most use cases.
Key Takeaways
- Manifest deprecated its LLM router after four months of mixed results, concluding that a single, reliable model is often more effective for most use cases.
- Complexity of tasks cannot be accurately determined from prompts alone, as additional context is often revealed through subsequent actions.
- Caching is significantly more cost-effective than model routing, with cache reads being 75% to 90% cheaper than uncached inputs.
- Switching between different models during workflows leads to lower quality work and unpredictability, which can complicate maintenance and management.
Community Sentiment
Positives
- Some commenters highlight the practicality of using cheaper models for personal tasks, emphasizing the importance of cost-effectiveness in LLM selection.
- There's a belief that defining subagent roles linked to specific models can enhance efficiency, with one user sharing their success in optimizing workflows.
Concerns
- Skepticism looms around the idea of LLM routing, with many arguing that understanding the subtleties of different models is impractical given the rapid pace of new releases.
- Commenters express doubts about the viability of routing as a long-term solution, suggesting that model providers are better positioned to solve these challenges themselves.
