OpenRouter launches Model Fusion compound system

While Model Fusion represents a significant step toward compound AI systems, its multi-model orchestration structure makes it a high-cost, high-latency overkill that is counterproductive for coding tasks.

* **Coherence vs. Consensus:** Coding requires maintaining a single, consistent state and architecture; mixing snippets from multiple models via a judge often leads to syntactically correct but structurally disjointed code.

* **The "Looping" Overhead:** Multi-model consensus and multi-turn checking loops naturally improve open-ended research reports but introduce a "feedback tax" and latency that drag down iterative coding tasks.

* **Pricing and Latency Penalty:** Running several models in parallel plus a synthesis step dramatically inflates costs and latency, making it impractical compared to single frontier coding models like Claude 3.5 Sonnet.

* **Misleading Benchmarks:** Generalized "Deep Research" benchmarks that benefit from more loop iterations fail to translate into correct software engineering capabilities in real-world codebases.