A Well-Scored Model Can Still Win Zero Head-to-Heads
A Well-Scored Model Can Still Win Zero Head-to-Heads
On summarizing structured content, GLM-5.3 scores 8.41 (±0.24) after 34 independent judgments, tied with Grok 4.7 [^1]. Across 40 occasions where judges compared it with rivals on the same work, it was never picked as the best answer [^1]. Both can be true at once, because they measure different things.
Two numbers, two questions
A mean score says how good a model’s answers are on a 10-point scale. A win share says how often its answer was judged the best among the rivals in the same contest. The second is winner-take-all: a close second counts the same as last place.
That makes win share harsh in a field of near-equals. On summarizing content, the top two answers finished within half a point in 89.6% of 144 judged occasions, with a median gap of 0.2 [^2]. A model that is reliably good but rarely on top can score well and still collect no wins. That is an interpretation of how the metric behaves. The results do not say where the zero-win models placed.
Qwen 3.7 Plus shows how far apart the two numbers can sit. It was judged best in none of 78 occasions on summarizing content [^2]. It is also one of the better-measured models in the category, at 8.26 (±0.16) after 58 judgments on weighing several perspectives into a decision [^3] and 7.96 (±0.17) after 107 on structured summaries [^1]. Those scores come from other tasks, so they cannot explain the zero. They do show that a no-win record can sit beside solid, well-measured scores.
Who was never preferred, and who won most
Several models went without a single win on summarizing content: Claude Sonnet 4.6 across 91 occasions, DeepSeek V4 Pro across 96, Claude Sonnet 5 across 80 and GPT-5.6 Terra across 53 [^2]. On structured summaries, Gemini 3.8 Flash had none across 40 [^1].
At the other end, nobody dominates. Even the best shares reported are under half:
| Model | Job | Judged best | Share |
|---|---|---|---|
| Thinking Machines Inkling | Summarizing content [^2] | 15 of 34 | 44.1% |
| Claude Opus 5 | Weighing perspectives into a decision [^3] | 14 of 36 | 38.9% |
| Meta Muse Spark 1.1 | Writing a synthesis analysis [^4] | 12 of 34 | 35.3% |
| GPT-5.6 Sol | Summarizing content [^2] | 10 of 30 | 33.3% |
| NVIDIA Nemotron-3 Ultra 550B | Writing a synthesis analysis [^4] | 11 of 34 | 32.4% |
| Moonshot Kimi K3 | Writing executive summaries [^5] | 8 of 26 | 30.8% |
Winning a third or so of contests makes a model a frequent winner in a spread-out field. It does not make it a safe single choice for best-answer work.
When win share should change your pick
Which number to trust depends on what the step does with the output.
- The step keeps one answer. Choosing among several drafts, or sending a summary out unedited, rewards being best, not being good. Check win share alongside the score, and leave zero-win models off that shortlist.
- The step needs acceptable output every time. The mean score and its interval are the better guide. A no-win record is not a reason to drop a model here.
- Judge a no-win record per task. GLM-5.3, never preferred on structured summaries, leads executive summaries at 8.61 (±0.18) [^6].
What a win share cannot tell you
A win share depends on which rivals were in each contest. Shares from different fields cannot be ranked against each other, and none can be ranked against the 10-point scores. The results also do not say which rival won, how close each contest was, or what the models cost. A no-win record means judges did not prefer a model, not that its output is poor.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar you need and see which models clear it. For any step that keeps a single answer, also check how often each candidate was judged best on that task.