Collaboration Beats Capability: What OpenRouter's "fusion beats frontier" finding means for the PAICE portfolio, and the two open specs we are building for machine-to-machine collaboration

PAICE (People + AI Collaboration Effectiveness) rests on a claim that sounds almost too simple: how well someone collaborates with AI predicts outcomes better than how capable either party is alone. A terse professional who catches an injected error scores higher than a fluent one who misses it. Collaboration effectiveness beats raw capability.

Last week, OpenRouter published a finding that shows the same law holds one layer down, between the models themselves. They call it "fusion beats frontier." It is worth understanding, because it is the clearest external evidence yet that the thing PAICE measures in people is becoming the thing that decides outcomes in machines.

Watch the Video

The finding: fusion beats frontier

OpenRouter ran 100 deep-research tasks from Perplexity's DRACO benchmark, which scores reasoning, tool use, synthesis, and citation quality. They compared single frontier models against panels: send the same prompt to several models in parallel, then have a synthesizer model fuse the answers into one.

The single-model baselines were what you would expect. Claude Fable 5 led at 65.3%, GPT-5.5 reached 60.0%, Claude Opus 4.8 came in at 58.8%.

Then the panel of Fable 5 and GPT-5.5, synthesized by Opus 4.8, scored 69.0%. The combination beat every individual model tested, including the strongest one running alone.

Two further results matter more than the headline.

A budget panel of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro scored 64.7%, within a point of Fable 5 running solo, at roughly half the cost of the frontier panel. Cheaper models, arranged well, nearly matched the best single model in the world.

And the cleanest result of all: a panel of Opus 4.8 fused with a second copy of Opus 4.8, judged by Opus 4.8, scored 65.5%. That is the same model, run twice, fused, beating its own solo score of 58.8% by 6.7 points. The gain did not come from mixing different architectures. A meaningful share of it came from the act of structured combination itself.

The scope is honest and narrow. This is one benchmark, focused on deep research. It does not claim a panel beats a frontier model at everything. But within that scope, the direction is unambiguous: structure applied across models outperformed capability concentrated in one.

The PAICE thesis, one layer down

Read that last result again. Holding the model constant and changing only the collaboration structure produced most of the lift. That is the experiment PAICE has been running on people the whole time.

PAICE does not measure how smart you are. It measures what happens in the space between you and the AI: whether you verify, whether you catch what slips through, whether you hold accountability for the output instead of laundering it through a tool. The premise is that this space, not the raw horsepower on either side, is where results are won or lost. Fusion beats frontier is that premise, demonstrated between models, with numbers.

This is also the argument our Signals and Subtractions newsletter made about teams: the limiting factor is rarely model capability, it is collaboration design. The same law now shows up at three layers. Within a team of people. Between a person and an AI. And between the models themselves. At every layer, the system that collaborates well beats the strongest part working alone.

It is also why PAICE has always been model-agnostic by design. If outcomes were decided by which single model you picked, the rational move would be to bet everything on the current leader and re-platform every time the leaderboard changes. Fusion beats frontier says the opposite. The durable advantage is in how the pieces are arranged, audited, and governed. That is portable across models. The leaderboard is not.

If collaboration between models is where the gains are, then collaboration between models needs infrastructure. Two open specifications in the PAICE portfolio are being built for exactly that. One gives them rules of engagement. The other gives them a language.

Collaboration needs rules of engagement: Turnfile

When peer agents disagree, something has to govern how the disagreement resolves. That's what we've been openly working with via our Turnfile protocol since February.

Most multi-agent systems assume a boss. One model plans, the others execute, and disagreement is hidden because subordinates do not get to object. Turnfile inverts that. Agents work as peers in owned lanes, counter-recommendations are first-class, disagreement surfaces before action rather than after, and a human stays on the loop as arbiter rather than copy-paste relay. Every decision lands in plain markdown that you can reconstruct without any special tooling.

The fusion result gives Turnfile its sharpest framing yet. A panel only beats the frontier if the models genuinely disagree and the synthesis step does real work. A room of yes-men averages out to mediocrity. Turnfile is the protocol for making disagreement productive instead of suppressed, and for keeping the whole negotiation auditable while it happens.

It is no longer a thought experiment. Turnfile now runs live sessions with three different model families negotiating as peers, Claude, Codex, and Gemini, reaching consensus across owned lanes with a human holding intent and veto. The protocol itself was built this way, by agents collaborating under it, which is the most demanding eval we could give it.

Collaboration needs a language: Tokenese

Rules of engagement still need a channel to run on. The newest addition to the portfolio is Tokenese, a token-native interlingua for LLM-to-LLM communication.

When a panel of models works on a problem, they talk to each other in English, a language shaped by human constraints: serial speech, social hedging, redundancy against a noisy room. Models inherit all of that overhead and get none of the benefit. Tokenese is a designed language for the machine-to-machine channel that is both denser and more precise than English, measured in real tokenizer tokens rather than characters.

A few design choices matter for a regulated reader:

It stays in token space. Plain text crosses the wire and each party tokenizes it independently. No shared embeddings, no latent channel, no vendor lock. A Tokenese exchange between two models from two vendors is still just text you can read off the wire.

Its vocabulary is audited, not asserted. A symbol enters the lexicon only if it has been measured to cost what it claims across every certified tokenizer, and the audit scripts ship with the spec so anyone can reproduce the claim. This is the same discipline PAICE applies to scoring: measured, not assumed.

It is human-auditable on purpose. A competent person with the one-page audit card can follow any conforming transcript. A dense machine language that produces opacity a human cannot unwind is rejected by the spec, regardless of how many tokens it saves. For anyone who has to explain to a regulator what their AI systems said to each other, that invariant is the whole point.

Tokenese is early. The grammar is at v0.3 and the tooling passes its full test suite, with a live A/B measurement between models as the open question that decides whether the compression survives contact with real misparse rates. The bet is specific and falsifiable: that natural language sits so far from the efficient frontier for this channel that a designed language can win on density and accuracy at once.

Why a regulated professional should care

This can read as portfolio housekeeping. It is not. It is a forecast about your stack.

Multi-model is arriving whether or not you chose it. The moment a panel of cheaper models can match a frontier model at half the cost, every serious AI deployment starts routing across several models instead of standardizing on one. When that happens, three questions stop being optional.

Can you read what the models said to each other? That is an auditability problem, and it is why Tokenese treats human-readability as a hard invariant rather than a feature.

Can you govern how they resolve disagreement? That is a process problem, and it is what Turnfile makes explicit and recoverable in plain text.

Can your people supervise a system with several models in it instead of one? That is a collaboration problem, and it is exactly what PAICE measures. A professional who could verify one AI's output now has to hold accountability for a negotiated answer assembled from several. The skill that matters is not running a bigger model. It is governing a collaboration.

Fusion beats frontier is good news for cost and quality. It is also a quiet raise in the bar for oversight. The PAICE Portfolio is being built so that the raise is one you can actually meet.

Collaboration Beats Capability