Fusion at a glance

What to consider

• Multiple models agreeing is not independent verification; they can share training data, source errors, assumptions, and blind spots.
• The judge can omit a valid minority answer, mischaracterize disagreement, or synthesize a polished but incorrect conclusion.
• Fusion does not guarantee better results than a carefully prompted single model on a particular task.
• OpenRouter's launch evidence covered one 100-task, English, text-only deep-research benchmark, not every coding, agent, creative, or long-horizon workflow.
• OpenRouter changed the benchmark judge and made other implementation choices, so its scores are not directly comparable with the original DRACO paper.
• One reported model completed only 93 of 100 tasks because of content filters, making some headline comparisons uneven.
• When invoked, Fusion is documented as often taking two to three times longer than a standard call.
• A default three-model panel is estimated at four to five times the cost of a single completion, and larger or reasoning-heavy panels can cost much more.
• The router's model page may display a zero top-level alias price, but Fusion documentation says users pay the sum of underlying model calls; budget from actual usage records.
• The outer model may decide not to invoke Fusion unless tool choice is forced, so merely selecting the slug does not prove panel deliberation occurred.
• Forcing Fusion on every request wastes latency and spend on greetings, classification, simple extraction, and other routine work.
• Latest-model aliases can change the concrete models and behavior without an application deployment; pin versions when reproducibility matters.
• Context limits depend on selected models, despite any large context figure displayed for the router alias.
• The prompt and retrieved material pass through multiple models and potentially multiple providers, expanding privacy, residency, contractual, and security review.
• OpenRouter itself does not retain prompt content by default, but provider data policies vary; enforce approved provider, no-training, and ZDR constraints where required.
• Web search and fetch can retrieve low-quality, manipulated, copyrighted, malicious, or prompt-injecting content that every panel member then treats as context.
• Search citations can be broken, secondary, outdated, or unrelated; a judge comparing them does not replace opening and validating the sources.
• Model refusals, outages, rate limits, tool failures, truncated outputs, and changing provider availability can alter panel coverage.
• High-stakes medical, legal, financial, employment, safety, or security conclusions still require qualified humans and authoritative primary evidence.
• A benchmark improvement does not establish ROI; teams need task-specific evaluation against accuracy, latency, cost, source quality, and human-review time.