OpenRouter launches Fusion API to mix outputs from multiple LLMs, claiming Fable-level performance at half the cost

neat thing I found months ago: if you take the union of terminal bench 2.0 tasks solved by opus 4.6, gpt 5.4, and gemini 3 pro, you get 92% vs. 75% for the leading model at the time

@jerryjliu0 This is basically the idea behind something I open-sourced. A bunch of people wanted to actually try model fusion in their Claude Code, so I built gavel. It runs Claude + Codex + Gemini on the same task in parallel and has Claude fuse their answers. https://github.com/junkim100/gavel

@jerryjliu0 I tested it and it does not give the same results when used for coding

I used MiMo-v2.5, Deepseek-v4-flash and Qwen3.7-plus and the combination did worse than Qwen3.7-plus standalone

Also note that it was significantly more expensive

@jerryjliu0 You don't even need a router. Simply have a second LLM review and critique the plan from the first LLM. This approach frequently leads to major improvements, helps avoid problems, and reduces the total number of reasoning cycles.

@jerryjliu0 I am taking this release with a massive grain of salt. Do you really think OpenAI, Anthropic, even Meta / xAI would miss something so obvious? Not to mention the benchmarks not passing the smell test.

@jerryjliu0 这不是骗子是什么？

@jerryjliu0 We have been doing multi-model code reviews that have significantly outperformed anything available on the market. This release from OpenRouter confirms the intuition.

@jerryjliu0 Labs built the models and gave away the orchestration advantage. Classic.

@jerryjliu0 Yeah, this. Most of the work is routine: summaries, retrieval, basic rewrites. You don't need top-of-the-line for that.

The economics flip the moment you route by task type.

@jerryjliu0 I too was like "What?"! Great thinking on the side of the @OpenRouter Yet I am going to try at local level!

@jerryjliu0 but it also shows deepseek is better than gpt-5.5 🙄🤥

@jerryjliu0 size doesn't mean shit anymore

@jerryjliu0 huge if true, compound models feel very underrated

@jerryjliu0 Pretty much. That's what Documind does (http://www.documind.cloud) to process documents. No need to settle for a single LLM, other can cost, which is bound to be diminished in the future plus can be optimized for different verticals.

@jerryjliu0 The Pareto frontier argument is sharp. The problem is compound systems have compounding failure modes, so you're trading model predictability for ensemble accuracy and that trade-off bites hard in regulated use cases.

@jerryjliu0 One thing not many people are talking about is how this approach is an absolute cache buster

@jerryjliu0 三个臭皮匠是顶不了一个诸葛亮的

This matches what shows up in production. For most knowledge-work tasks a routed mix of smaller models clears the bar at a fraction of the cost, and the frontier model only earns its price on the genuinely hard steps. The skill is knowing which step is which. Treating one model as the answer to every task is how budgets quietly blow up.

@jerryjliu0 如何进行编排的？

@jerryjliu0 Huh? You guys are awesome but the chart makes no sense