AI HOT (Curated Pool)Jun 12, 2026, 08:00 UTC+8

OpenRouter's model fusion panel beats GPT-5.5 and Claude Opus 4.8 on deep research benchmark

OpenRouter融合预算模型面板超越GPT-5.5和Claude Opus 4.8

OpenRouter launched Fusion, which sends a prompt to multiple models in parallel and has a judge model synthesize the final answer. On 100 DRACO deep research tasks, Fable 5 + GPT-5.5 fused scored 69.0%, beating Fable 5 alone at 65.3%. A budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro hit 64.7%—close to Fable 5 at roughly half the cost. The post doesn't disclose added latency or the exact per-call price for the budget panel.

Why it matters: OpenRouter's Fusion lets budget model panels beat solo frontier models on deep research via multi-model deliberation + judge. Concrete DRACO benchmark data and anti-cheat design make it worth reading. Score capped at 78 because it's a platform feature launch, not a model break...

Read the original ↗Export Markdown

Take

The reason to click: OpenRouter turned model ensembling into a product. You pick a panel of models and a judge, Fusion fires the same prompt to all of them in parallel, then the judge synthesizes one answer. On 100 DRACO deep research tasks, Fable 5 + GPT-5.5 fused hit 69.0%, beating Fable 5 solo at 65.3%. The budget panel—Gemini 3 Flash, Kimi K2.6, DeepSeek V4 Pro—scored 64.7%, close to Fable 5 at roughly half the cost.

I'd discount this on two fronts. First, the test set is only 100 tasks, and Fable 5's content filters blocked 7 of them, so the sample is even smaller. Second, the post says nothing about latency. Calling multiple models and waiting for a judge to synthesize will be slower than a single call—that's a real product constraint. The judge model (Opus 4.8) also adds cost and potential bias, neither of which is discussed.

Don't read this as "ensembles always win." The more useful take: on deep research tasks that mix reasoning, tool use, and knowledge retrieval, different models miss different things, and a fusion step can catch what individuals drop. But you're paying for multiple API calls plus waiting time. Worth trying if your task is latency-tolerant and accuracy-sensitive.

Source: AI HOT (Curated Pool) · openrouter.ai
Published: Jun 12, 2026, 08:00 UTC+8 · Jun 12, 2026