For three years the conversation about AI capability has been a leaderboard of single models. One name on top, a race to replace it, and policy debates organized around who controls that one name. A result published this month suggests that framing is incomplete.
What was published
On June 12, 2026, OpenRouter published benchmark results for Fusion, a feature that sends one prompt to several models at once, has a judging step compare their answers, and then writes one combined answer. They tested it on DRACO, a set of 100 deep-research tasks.
The numbers, as OpenRouter reported them:
- A panel of two models, combined by a third, scored 69.0 percent.
- The stronger of those two models, working alone, scored 65.3 percent. The other scored 60.0.
- A budget panel of three smaller models scored 64.7 percent, within a point of the best single model, at about half the cost.
- One model run twice and combined scored 65.5 percent, against 58.8 percent for the same model run once.
These are the vendor's own figures from one benchmark, and they come with caveats worth reading. The best single model was scored on 93 of the 100 tasks, because some tasks were blocked by its content filters. The benchmark is text-only and in English. And a combined answer takes two to three times longer than a single call. Treat it as one strong data point, not a settled law.
Why a panel can beat its members
A model is a fixed thing. It has a training cutoff, habits, and blind spots that come from its data and its makers. On a hard question, one model gives you one line of reasoning.
A panel gives you several. Each model finds some things the others miss, and some models make confident mistakes the others do not. The judging step is where the value is created: it finds what the answers agree on, flags where they contradict each other, keeps what only one model found if it holds up, and notes what nobody covered.
The same-model result is the interesting one. Two runs of one model, compared and combined, gained almost seven points. Much of the lift comes from the comparing and judging, before you mix vendors at all.
Two design choices matter a great deal. The panelists answer independently, without seeing each other. And the judge sees the answers without names, so it cannot favor a model from its own family. Try both below.
What this means when a model is taken away
Here is the part that matters for government teams and regulated industries.
Organizations restrict models. An agency can take a model off its approved list. A contract can rule out a vendor. A security review can bar a set of weights. Those are real controls, and they work on a single named model.
A system built around one model breaks when that model is removed. A system built as a panel loses one input and keeps working with the models that remain. The answer may be weaker, and the strongest panels still use strong models. But the work does not stop, and nothing has to be rebuilt.
We are not claiming that the best model no longer matters, or that access controls are pointless. The narrower claim is the useful one: depending on any single model is a fragility, and a panel reduces it.
Questions to ask about any system you buy
If your team relies on AI, or is about to, these questions apply whether or not the system uses a panel:
- If one model is removed from our approved list tomorrow, what breaks? You want an answer measured in configuration changes.
- Does the record show which model produced each answer? If a model is later restricted, you need to find the work it touched.
- Who judges the answers, and can the judge see which model wrote them? A judge that knows the authors can favor its own kind.
- What does the extra quality cost in time? Combining answers is slower. Know where that trade is worth it.
How we use it
We run a panel like this internally to check our own hardest calls. The panelists answer independently and the judge sees them anonymized. What we noticed matched the published results: the combined answer was steadier and less often confidently wrong than any single model we put into it.
For the record-keeping side of the question, see Why SAIN is building the trust infrastructure for AI. If your team is planning for a list of allowed models that can change, talk to us.