On June 12, 2026, OpenRouter launched Fusion: a single API call that fans your prompt out to a panel of models in parallel, each running with web search and web fetch tools; a judge model then reads every panel response and produces a structured analysis — consensus points, contradictions, partial coverage, unique insights, blind spots — which the calling model uses to write the final answer. The whole pipeline runs server-side and behaves like an ordinary model call.
The headline is the benchmark: on the DRACO deep-research benchmark, fused panels outscored every solo model tested. And because Fusion ships as a model slug, existing OpenRouter integrations can try it by changing one string.
How Fusion Works
Three layers. The panel: your chosen models each execute the full task with openrouter:web_search, openrouter:web_fetch (via Exa), and openrouter:bash. The judge: it reads all panel outputs and maps where they agree, contradict, and partially cover the question. The synthesis: the calling model writes the final answer from that analysis. In other words, Fusion is not a blend of outputs — it treats multiple models as independent retrieval and argument sources, then collapses judgment into a single exit point. Since panel members run with identical tools, their disagreements are evidence rather than noise: two models that browsed the same sources and still contradict each other tells you something no single model will volunteer.
On DRACO, Fused Panels Beat Every Solo Model
DRACO (arXiv 2602.11685, from Perplexity AI) contains 100 deep-research tasks across 10 domains, graded against roughly 39 weighted rubric criteria in four categories — factual accuracy, breadth and depth, presentation quality, citation quality — with each criterion scored three times. OpenRouter used Gemini 3.1 Pro Preview as judge, sanity-checked with Claude Sonnet 4.6.
The results: the top configuration, Claude Fable 5 + GPT-5.5 with Opus 4.8 synthesizing, scored 69.0%; a three-model panel of Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro hit 68.3%. The best solo model was Claude Fable 5 at 65.3% — and even that came from only 93 of 100 tasks, because its content filters refused 7 (for what happened to that model next, see our earlier coverage). The budget story is more interesting: Gemini 3 Flash + Kimi K2.6 + DeepSeek V4 Pro, synthesized by Opus 4.8, scored 64.7% at roughly half of Fable 5’s cost — within 1 point of the frontier, and 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 managed just 45.4%. One caveat for scoreboard watchers: OpenRouter’s judge differs from the paper’s, so these scores are not comparable with the original DRACO leaderboard.
Self-Fusion and a Clean Contamination Control
Two details worth remembering. First, fusing Opus 4.8 with itself — two instances, one synthesis — scored 65.5% versus 58.8% solo, a 6.7-point lift. A large chunk of Fusion’s gain comes from the synthesis step itself, not just model diversity. Second, mid-testing, panel models actually found the DRACO rubric online through web search. OpenRouter blocked the sources with excluded_domains on web_search and blocked_domains on web_fetch, and states that every published number is post-exclusion.
Four Ways to Call It, and the Latency Bill
Integration options: a chatroom at openrouter.ai/fusion; the model slug “openrouter/fusion”, which injects a default frontier panel; a server tool — add openrouter:fusion to your tools array and let the base model decide when to invoke it — or a plugin with a custom panel and synthesizer. The server-tool route is the one to watch: it turns Fusion into a capability the base model can pull in when a question deserves a panel, instead of a routing decision the developer hard-codes. The cost is latency: Fusion-invoked calls run “often 2-3x longer than a standard call” because the pipeline waits for every panel member, and OpenRouter published no dollar pricing — only relative cost comparisons. The limitations section is unusually candid: DRACO is text-only, English-only, and static; absolute scores swing 10–25 points with the judge model, though rankings hold; long-horizon tasks were not tested at all — exactly where Fable 5 shines; and it is not a coding-model replacement. The recommended pattern is a coding model calling Fusion selectively, at decision points like architecture choices.
Sources
- Surpassing Frontier Performance with Fusion — OpenRouter
- openrouter/fusion model page — OpenRouter
- Openrouter Fusion API — Hacker News
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.