DecisionNode MoE reached 81.6% mean accuracy across 15 text suites in Bynn’s preliminary October 2026 evaluation. The suites contain 25,134 items and cover a range of decision tasks. The published comparison includes results where other models lead, and this article walks through both: what was measured, how, where the model is strong, where it trails, and how an enterprise team should read a benchmark like this before building on it.
DecisionNode MoE: 81.6% mean accuracy across 15 text suites in Bynn’s preliminary October 2026 evaluation
What Bynn’s preliminary, Bynn-run evaluation of DecisionNode measured across 15 text suites and 25,134 items, where the model leads and trails, and how an enterprise team should read an AI decision engine benchmark.
- Bynn Intelligence
- 5 min read
DecisionNode MoE81.6%mean accuracy15 text suites, 25,134 items81.6050100
A preliminary, Bynn-run evaluation
These results come from an evaluation Bynn ran itself, on the development splits of each suite. It is not an independent result, it is distinct from an independent evaluation such as NIST’s, and the figures may change when the full run is published.
What DecisionNode is
DecisionNode is Bynn’s proprietary System One decision model and inference service for workflows, applications, and robotics. It interprets text, images, and structured inputs, then returns decisions that connected systems can act on. Where a chat model writes an answer token by token, a decision engine returns a typed answer, a choice, a score or a truth value, with a confidence your code can branch on. That is the kind of component an enterprise AI system needs at the point where something has to happen next.
What was measured
The evaluation covers 15 text suites with 25,134 items in total. Each suite tests one kind of decision: applying a written policy to a case, picking the right element or action on a web page, holding to the task when the input tries to talk the model into a wrong answer, choosing among dozens of options, finding the answer deep inside a long document, answering the same decision in many languages, and answering hard multiple-choice questions that need broad world knowledge.
Each suite is scored as accuracy: the share of items where the answer matches the reference. The headline figure is the mean of the 15 suite accuracies. DecisionNode MoE scores 0.816. DecisionNode Flash, the tier built for speed and price, scores 0.784.
How it was measured
- Development splits of each suite, October 2026.
- 95% intervals for Bynn’s models and the closest comparison system, from a bootstrap that resamples whole inputs.
- The commercial comparison system was called through its official API with its documented request shape. Open models ran on Bynn’s GPUs, served the way their authors publish them.
- Nine systems appear in the mean comparison, Bynn’s two models among them.
Where it leads and where it trails
A mean hides the shape of a result. The system with the next-highest mean in the comparison scored 0.803, so the difference in the mean is 1.3 points; the differences on individual suites are much larger, in both directions. The figure below shows DecisionNode MoE minus that system on every suite.
DecisionNode MoE is clearly ahead, with the two 95% intervals apart, on three suites: prompt injection (+16.4 points), web tasks (+6.7) and rules and policies (+6.0). That is the everyday work of a workflow: hold to the task when the input argues, act on a page, apply a written policy.
It is clearly behind on two: knowledge (MMLU-Pro), by 12.2 points, and new task families, by 2.6 points. On the remaining ten suites the difference is inside the noise. If your decisions lean on broad world knowledge, weigh the knowledge suite first.
| Suite | Items | DecisionNode MoE | Difference, points |
|---|---|---|---|
| Prompt injection | 798 | 0.764 | +16.4 |
| Web tasks | 1,106 | 0.765 | +6.7 |
| Rules and policies | 1,759 | 0.891 | +6.0 |
| Large option sets | 1,677 | 0.806 | +3.1 |
| Public decision set | 117 | 0.880 | +2.5 |
| Familiar task families | 4,100 | 0.796 | +2.2 |
| Short classifications | 178 | 0.921 | +1.7 |
| Multilingual | 1,812 | 0.796 | +1.4 |
| Open decision tasks | 5,011 | 0.795 | -0.3 |
| Held-out decisions | 636 | 0.940 | -0.8 |
| Transfer tasks | 730 | 0.838 | -1.4 |
| Long inputs | 1,147 | 0.835 | -1.8 |
| Typed decisions | 955 | 0.713 | -2.2 |
| New task families | 3,629 | 0.797 | -2.6 |
| Knowledge (MMLU-Pro) | 1,479 | 0.698 | -12.2 |
The difference column compares DecisionNode MoE with the system that has the next-highest mean. The published comparison names every system measured and shows each suite with every per-suite result it publishes.
Speed, measured separately
In the same preliminary October 2026 measurement, DecisionNode Flash returned a short decision in 19 ms end to end, measured from one client with the network included; about 5 ms of that was spent on the server. That is one short decision on one model from a single client, not a latency guarantee for other inputs, models or networks. DecisionNode MoE’s latency is not yet published.
What it means for an enterprise buyer
A decision engine earns its place inside a process when two things hold: its answers are right often enough for the decision at hand, and its confidence tells you when they are not. The suite-by-suite view answers the first question better than any single mean.
- Match the suite to the decision. A claims-routing or policy-application workflow resembles the rules and policies suite more than the knowledge suite. Start from the suites closest to your workload.
- Treat the mean as a summary. 81.6% across 15 suites is not a promise of 81.6% on your task. It can be higher or lower.
- Test on your own data. With a free account, the DecisionNode playground runs any request against both models, and the same request gives the same answer every time, so a test is repeatable.
- Plan for low confidence. Set a threshold, act on answers above it, and send the rest to a safer path or a person.
A simple example shows the role. A package reaches a sorting line. The camera sees it. DecisionNode evaluates its condition and label, selects a route, and returns a decision the connected workflow can execute. The same approach extends to software workflows, including checking an AI agent’s proposed actions before they run.
Limits and caveats
- Preliminary and Bynn-run: measured on development splits, not an independent evaluation.
- A mean across suites, not a guarantee of accuracy on any single task.
- Calibration results per question type will be published with the full run; this article does not report them.
- Every system in the comparison changes over time, DecisionNode included. The scores describe the versions measured in October 2026.
Read the full comparison
The methodology, every suite and every system are published at decisionnode.com/benchmarks. For what DecisionNode does and how it connects to workflows and robotics, see the DecisionNode page. Our results, each labelled by type of evidence, are on the research page, and if you want a decision engine inside your own process, see how we approach enterprise AI.
