When Averages Tie, Count the Wins
When Averages Tie, Count the Wins
On the Fronset benchmark’s social-content selection tasks, average quality scores bunch so tightly that they often cannot tell models apart. Head-to-head results, which record how often a model’s answer is judged the best of several answers to the same input, can separate models where the scores cannot. They also expose models that are never judged best at all, which an average score can hide.
On selection tasks, the scores tie
Post selection, choosing which posts to publish together, is the most crowded area of the category. On the 10-point scale, Gemini 3.8 Flash and Meta Muse Spark 1.3 score 8.73, ±0.18, GPT-6 Sol 8.7, ±0.2, and Claude Opus 5.5 8.6, ±0.21 [^1]. Their ranges overlap, which this article, like its source, treats as a tie: quality alone does not choose between these models.
The tie holds answer by answer, too. A judged occasion is one live input on which several models’ answers were scored. On post selection, the top two answers scored exactly the same in 70.0% of 60 judged occasions, with a median gap of 0.0 [^1], and on choosing an X post the top two tied exactly in 64.3% of 84 judged occasions [^2].
Head-to-head wins add what the scores miss
A win share is the fraction of judged occasions on which a model’s answer was judged best. Moonshot Kimi K3 was judged best in 15 of 25 occasions (60.0%) when choosing an X post [^2], and Claude Opus 5.5 in 10 of 28 occasions (35.7%) on portfolio selection [^1]. These are win shares against other models answering the same live input, not scores, and the source does not give the other models’ shares, so they cannot be ranked against the score tables.
Deciding which subreddit a post belongs in is the clearest exception to the tie. There the top two answers finished within half a point in only 48.6% of judged occasions, with a median gap of 0.6 [^3]. gpt-5.6-sol was judged best in 25 of 29 occasions (86.2%), against gpt-5.6-terra at 17 of 41 occasions (41.5%) on the same task [^3], and the source reads gpt-5.6-sol as the stronger of the two.
One caution applies to every win share here. The shares rest on differing numbers of occasions, 29 and 41 in the subreddit case, and no confidence intervals are given for any of them.
Some models are never judged best
The starkest head-to-head result is a model that never wins. On engagement review, kimi-k2.6 was chosen best in none of 41 occasions, and deepseek-v4-pro in none of 41 [^4]. For deepseek-v4-pro the findings note that an average quality score does not reveal this pattern [^4]. On campaign generation, NVIDIA Nemotron 3.5 Lightning was never chosen best in any judged occasion [^5].
The source qualifies this carefully. Never being preferred does not say how close a model came, so it is not proof of unacceptable output. It is not a basis for routing to these models either.
Why the two measures disagree
An average score grades each answer on its own; a win share asks which answer came out on top against the others on the same input. When answers bunch as tightly as they do on post selection, the two can come apart. A model that is consistently a little behind the best answer can post a respectable average while winning rarely or never.
The source treats scores and win shares as separate measures, and no model leads on all of the measures it reports. The score says how good an answer was; the win share says how often it was the best one available. Reading only the first hides the second.
What to do with it
On the selection tasks, do not break a score tie on the second decimal. Look at the head-to-head results where the benchmark gives them, and test the tied models on your own posts before committing.
Where the head-to-head result is lopsided, let it decide. On subreddit choice, the source routes to gpt-5.6-sol as the stronger of the two measured models [^3].
Do not route a task to a model that is never judged best there on the strength of its average score. That covers kimi-k2.6 and deepseek-v4-pro on engagement review [^4] and NVIDIA Nemotron 3.5 Lightning on campaign generation [^5].
The measurement behind these figures is continuous and ongoing, so check the live figures at https://fronset.ai/benchmark/ before you settle on a model.