Cluster-Labeling Scores Do Not Predict Theme Discovery

The model that names your topic clusters best is not necessarily the one that should find them. On the Fronset benchmark, every model measured on both steps scores lower at discovering themes in text than at labeling clusters it is handed, and some models tied near the top of labeling are among the weakest at discovery [^1][^2].

Two steps that look alike

Labeling gives a model a group of related content and asks for a name. Theme discovery gives it raw text and asks it to decide what the groups are. Both fall under topic organization, so it is tempting to pick one model for the whole pipeline from whichever leaderboard you check first.

If that is the labeling leaderboard, it will tell you little, because it barely separates models. Moonshot Kimi K3 has the highest figure, 8.87 out of 10, but GLM-5.3 at 8.69 and GPT-6 Sol at 8.56 have overlapping intervals and are tied with it [^1]. The whole reported field sits between 8.01 and 8.87 [^1].

One reading of the gap, offered as interpretation rather than something the benchmark measures: naming a structure that already exists is close to summarizing, while discovery makes the model choose the structure itself. The figures fit that picture but do not test the mechanism.

How far discovery scores fall

These models have results on both steps:

The drop is uneven. GLM-5.3 loses less than a point. GPT-6 Sol loses exactly two, and GPT-6 Luna more than two. Claude Opus 5.5 posts the same labeling score as GLM-5.3 but finishes more than a point below it on discovery.

The models that hold up best are GLM-5.3 and its Flash variant. GLM-5.3 Flash has the highest discovery figure, 8.15, and the two are tied within their intervals [^2].

So a labeling leaderboard that looks like a flat tie hides a real spread one step earlier in the pipeline. Choosing GPT-6 Luna or GPT-6 Sol for both steps because they rank well on labeling would put two of the weakest discovery models in front of your data.

What to do with it

Treat labeling and discovery as separate routing decisions:

• Discovery: shortlist on discovery results. GLM-5.3 and GLM-5.3 Flash are the current front-runners. Test them on your own inputs and keep a review step behind whatever they produce.
• Labeling: since the top group is tied on quality, choose on answer rate, latency and price. GPT-6 Luna, weak at discovery, is a reasonable labeler and the only one with a reported timing figure: it answers on 88% of attempts, with a median latency of 4.0 seconds over 26 synchronous calls [^1].

The discovery front-runner is also slow. GLM-5.3 has a median response time of 176 seconds on discovery, over 31 synchronous calls under a 600-second deadline [^2]. That belongs in a background job, not a user-facing request.

The caveat: answer rates and wide intervals

Scores are measured only on calls that returned an answer [^1]. GLM-5.3 answers on 76% of discovery attempts and GLM-5.3 Flash on 80% [^2], so a pipeline built on either needs retries or a fallback, and the score says nothing about the calls that returned no answer.

GLM-5.3 Flash’s lead is also thin. Its interval, ±0.3, is the widest in the discovery results [^2] and wider than the ±0.2 the benchmark requires before it counts a model as ranked on a task [^3]. Treat the GLM pair as the best current bet rather than a settled winner. No cost figures accompany these results, so price could still reorder the shortlist.

The live benchmark at https://fronset.ai/benchmark/ has a page for each task. Check theme discovery on its own rather than inferring it from labeling, set the quality bar your pipeline needs, and compare pricing there before you commit a model to each step.