Clustering’s Hard Part Is Finding Topics, Not Naming Them

In the Fronset benchmark, naming topic clusters is a step where most models land close together; finding the themes in the first place is not. Every model with results on both steps scores lower on thematic topic discovery than on topic cluster labeling [^1][^2]. And on discovery, the leading group includes routes at $0.04 and $0.45 per task run with tied scores, so paying more did not buy better themes [^2].

Labeling bunches together; discovery drops

Scores here are out of 10 with the ± range the benchmark reports; where two ranges overlap, treat the models as tied. On topic cluster labeling, most measured models reach the 8s with overlapping ranges. The head-to-head measurements, which compare the top two answers on each judged occasion, agree: those answers finished within half a point on 93.8% of 65 occasions, with a median gap of 0.2 [^1].

Thematic topic discovery, finding the themes across a body of text, is where those figures fall. Moonshot Kimi K3 goes from 8.91 on labeling to 7.65 (±0.35) on discovery, Claude Opus 5.5 from 8.61 to 6.74 (±0.26), GPT-6 Sol from 8.5 to 6.45 (±0.37) and GPT-6 Luna from 8.32 to 6.22 (±0.41) [^1][^2].

On discovery, price and score come apart

Discovery’s cost per task run varies widely, and higher price did not buy a better discovery result.

The two highest figures, Nemotron-3 Ultra 550B and GLM-5.3 Flash, overlap, so they tie, and both are among the cheaper routes here. GLM-5.3 costs far more than GLM-5.3 Flash and sits in the same leading group. Further down, Claude Opus 5.5 is the second most expensive route in the table and scores no better than DeepSeek V4.1 Flash, the cheapest.

One caveat applies to every price comparison. Cost bases are not uniform: some figures are measured on each model’s own usage and others are repriced at current rates, which makes cross-family cost comparisons approximate [^1]. The GLM-5.3 pair is not a cross-family comparison.

What this means for a clustering pipeline

On labeling, quality is a tie and price can decide. Discovery is where output quality is scarce, which makes it the step most likely to limit what a clustering pipeline produces. In these results, a bigger bill did not fix that.

Part of the discovery bill also buys nothing. In topic discovery and clustering, a related task, $10.26 of $74.20, a stated 13.8%, was billed for tokens that could not be parsed, though the material attributes the figure to no model [^3].

What to do about it

Start discovery with the tied leaders rather than the most expensive models. This guide’s route is GLM-5.3 Flash or Nemotron-3 Ultra 550B, and not GLM-5.3 on a tied score [^2].

Spend your evaluation effort on discovery, not labeling. On labeling, where the top answers are so often within half a point, a cheap model is a fair choice: GPT-6 Luna scores 8.32 at $0.0009 per task run in batch mode [^1]. On discovery, even the leaders sit near the 8.0 quality bar this guide uses, so test the themes on your own material before you rely on them.

When you set a budget for discovery, allow for unparseable output on top of the list price per run.

The task-level results behind these figures are on the live benchmark at https://fronset.ai/benchmark/, where you can set the quality bar your workflow needs and compare batch against sync pricing yourself.