Summary: OpenRouter seated Gemini 3.1 Pro Preview as judge, sanity-checked by Claude Sonnet 4.6. The bill is latency: a Fusion-triggered call runs “often 2-3x longer than a standard call” because the pipeline can’t finish until every panel member does, and OpenRouter has given no dollar pricing — nothing beyond relative cost comparisons. The limitations list is strikingly frank: DRACO is text-only, English-only, and static; swap the judge and absolute scores shift 10–25 points while rankings stay intact; long-horizon tasks went untested — exactly Fable 5’s home turf; and Fusion is no coding-model replacement.
Fusion went live on June 12, 2026. To the caller it is just one ordinary model call — the entire flow runs server-side. Inside, one prompt fans out in parallel to a panel of models, each completing the full task with web search and fetch tools; a judge model condenses all panel answers into a structured read-out — agreements, clashes, half-covered ground, one-model-only insights, model-specific blind spots — for the calling model to write the final answer from.
The pitch rests on the numbers: on DRACO, a deep-research benchmark, fused panels came out ahead of every solo model in the field. And since Fusion arrives as a model slug, an integration that already talks to OpenRouter can try it by editing one string.
Inside the Fusion Pipeline
The design has three layers. The panel: the models you pick, each running the whole task with openrouter:web_search, openrouter:web_fetch (via Exa), and openrouter:bash. The judge: reads every panel answer, charting where they agree, clash, or only partly cover the question. The synthesis: the calling model writes the final answer from that analysis. Fusion is no averaging of outputs: the models act as independent sources of retrieval and argument, and the verdict funnels through a single exit. Since every panel member carries the same tools, clashes read as evidence, not noise — two models that browsed the same web yet disagree expose what a single model never volunteers.
DRACO Results: Panels Sweep the Solo Field
DRACO (arXiv 2602.11685, from Perplexity AI) packs 100 deep-research tasks spread over 10 domains, graded on roughly 39 weighted rubric criteria grouped into four categories — factual accuracy, breadth and depth, presentation quality, citation quality — every criterion judged three separate times. OpenRouter seated Gemini 3.1 Pro Preview as judge, sanity-checked by Claude Sonnet 4.6.
The standings: Claude Fable 5 + GPT-5.5, synthesized by Opus 4.8, leads at 69.0%; the three-model panel of Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro follows at 68.3%. The solo record is Claude Fable 5’s 65.3% — earned by answering only 93 of the 100 tasks, its content filters declining the other 7 (where that model landed next: our earlier coverage). The budget panel is the quieter story: Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro, again with Opus 4.8 synthesizing, posted 64.7% for about half what Fable 5 costs — under 1 point off the frontier, ahead of solo GPT-5.5 (60.0%), solo DeepSeek V4 Pro (60.3%), and solo Opus 4.8 (58.8%); solo Gemini 3.1 Pro stalled at 45.4%. Caveat for leaderboard watchers: OpenRouter judged with a different model than the paper used, so these figures don’t line up with the paper’s DRACO leaderboard.
The Self-Fusion Signal, and Keeping the Test Clean
Two asides deserve a note. First, self-fusion: Opus 4.8 paired with a second instance of itself — two instances, one synthesis — scored 65.5% against 58.8% solo, a 6.7-point climb; much of what Fusion adds comes from the synthesis step itself, with model diversity only part of the story. Second, mid-run the panel models actually dug up the DRACO rubric through web search; OpenRouter shut that door with excluded_domains on web_search and blocked_domains on web_fetch, and every number it publishes comes from after that exclusion.
Four Integration Routes, and the Price in Latency
You can reach Fusion four ways: a browser chatroom at openrouter.ai/fusion; the model slug “openrouter/fusion”, which arrives with a default frontier panel attached; a server tool — drop openrouter:fusion into your tools array and leave the timing to the base model; or a plugin that takes your own panel and synthesizer. The server-tool shape is the one to watch: Fusion becomes a capability the base model summons when a question deserves a panel, not a routing decision a developer hard-codes. The bill is latency: a Fusion-triggered call runs “often 2-3x longer than a standard call” because the pipeline can’t finish until every panel member does, and OpenRouter has given no dollar pricing — nothing beyond relative cost comparisons. The limitations list is strikingly frank: DRACO is text-only, English-only, and static; swap the judge and absolute scores shift 10–25 points while rankings stay intact; long-horizon tasks went untested — exactly Fable 5’s home turf; and Fusion is no coding-model replacement. Official advice: let a coding model call Fusion selectively, at the decision points that matter — architecture choices, for instance.