OpenRouter's Fusion API turns a panel of models into a single call
OpenRouter just rolled out Fusion, an API feature that queries multiple models in parallel, adds web search, and synthesizes a final response using a judge model. In testing, a budget panel of three smaller models scored nearly on par with Anthropic's flagship Fable 5 at roughly half the cost. It effectively turns model diversity into a plug-and-play research team.
If you do serious research with LLMs, you probably know the drill: query one model, paste the prompt into another, maybe check a third, and spend five minutes manually piecing together the right answer.
On June 13, OpenRouter officially turned that entire workflow into a single endpoint with the full rollout of its Fusion API.
Turning models into a research crew
Fusion first surfaced as a public experiment back on March 31, 2026, and is now fully integrated into OpenRouter's production API. The premise is dead simple: instead of hunting for one perfect model, you spin up a small panel to figure it out together.
When you hit the openrouter/fusion alias, the system automatically routes your prompt across three to five models simultaneously. These models can run web searches and fetch pages in parallel to gather facts.
Once they draft their answers, a designated judge model compares the arguments, resolves contradictions, and delivers one synthesized response. You can stick with built-in presets like Quality or Budget, or manually pick the workers and the judge.
Half the bill, top-tier performance
The benchmark numbers explain why people are paying attention. In OpenRouter's DRACO test, fusing Anthropic's top-tier Fable 5 with GPT-5.5 scored 69.0%. That comfortably beat every single model tested, including Fable 5 on its own at 65.3% and GPT-5.5 at 60.0%.
Even more compelling is the budget setup. OpenRouter tested a panel combining Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro. It notched a 64.7% score, beating standalone runs of GPT-5.5 and Claude Opus 4.8 while landing within roughly one point of Fable 5.
In a June 12 evaluation across 100 complex reasoning tasks, Fusion again outperformed standalone GPT-5.5 and Claude Opus 4.8. Pricing is strictly pay-as-you-go: you pay the stacked token cost of all models invoked. Because fast budget models are so cheap, that three-model ensemble still runs at roughly half the cost of a single Fable 5 query.
Why it matters
For builders, Fusion eliminates the boilerplate of writing parallel scrapers, concurrent API dispatchers, and custom consensus prompts. If you already use OpenRouter, you can drop in this multi-model pattern without rewriting your backend.
It is also a shrewd business play for OpenRouter. The platform already aggregates over 400 models across more than 60 providers. By introducing a synthesis layer, OpenRouter turns cheaper commoditized models into competitive team players, softening the lock-in of single proprietary giants.
There is an obvious tradeoff: latency. Waiting for three to five models to search the web, generate answers, and then pass everything through an evaluator means this will feel sluggish for snappy conversational UIs. In our view, Fusion is not designed for chat; it is built for asynchronous deep-dive research, complex coding reviews, and background analysis where accuracy outweighs speed.
Sources (1)
- cryptobriefing.com(opens in a new tab)Originalcryptobriefing.com