The Model That Won Topic Discovery Lost the Next Step: Route Topic Pipelines by Sub-Task

If you are building topic clustering, labeling or taxonomy features on LLM APIs, the current benchmark evidence says there is no single model to pick. The same model that leads theme discovery collapses on content-domain suggestion, a model that clears the bar on taxonomy generation misses it on discovery, and the cheapest passing option is a different model again. The practical answer is to route each step of the pipeline separately, and to stop treating a win on one step as evidence for the next.

How the bar is set

The benchmark scores each model’s answers on a 10-point scale through judged comparisons against other models answering the same input, and reports a confidence interval alongside every score. For this guide, good enough means 8.0 out of 10 when the model answers.

Cost is reported per task run, tagged with the serving mode (sync or batch) and with whether it was measured on that model’s own usage or estimated because the cell has no qualifying usage history. That distinction matters below, because some of the leaders on individual steps are only estimated so far.

One model, two opposite verdicts

Start with the strongest discovery result in the set. GPT-5.6 Sol discovering themes across content scores 8.52 out of 10 (±0.32) and costs $0.51 per task run in sync mode, measured on this model’s own usage [^1][^2].

Now look at the same model on the adjacent step of proposing a content domain. GPT-5.6 Sol suggesting a content domain scores 3.84 out of 10 (±0.1) at $0.009 per task run in sync mode, measured on its own usage with repricing [^3][^4]. That is the sharpest negative in the whole category, and it comes from the model that leads discovery.

The step is not inherently hard; another model handles it well. Meta Muse Spark 1.1 suggesting content domains scores 8.71 out of 10 (±0.3) and $0.03 per task run in sync mode, measured on this model’s own usage and repriced at current rates [^3][^4]. A model family that is weak on one slice can lead another, and a high structuring score does not transfer to discovery.

It cuts the other way too

The flip is not a quirk of one vendor. Thinking Machines Inkling generating a taxonomy for publication sections scores 8.97 out of 10 (±0.12) and costs $0.01 per task run in sync mode, measured on this model’s own usage [^5][^6].

Send that same model upstream to discover themes across a collection and it drops below the bar, at 7.76 out of 10 (±0.49) and $0.09 per task run in sync mode, also measured [^1][^2]. A model can be a cheap, reliable taxonomy generator and still be the wrong choice for finding the themes that taxonomy will hold.

For reference, the highest score anywhere in the set is Moonshot Kimi K3 generating taxonomies for publication sections, at 9.33 out of 10 (±0.06) and $0.03 per task run in sync mode, measured on this model’s own usage [^5][^6]. Even that result says nothing about how Kimi K3 would do on discovery or suggestion.

Where the cheap seats are

At the routine end of the pipeline the numbers get very small. The cheapest measured option that still clears 8.0 is GLM-5.3 Flash assigning a section to a topic, which scores 8.16 out of 10 (±0.49) and costs $0.001 per task run in sync mode [^7].

That is not a like-for-like comparison with taxonomy generation, since the two tasks produce different outputs. It does mean that if your pipeline spends most of its calls on assignment rather than on building the taxonomy, the expensive model belongs on the rare step and the cheap one on the frequent step.

One step still has no firm answer. Matching topics to an existing taxonomy is estimated-only for every model compared, so the source recommends sending it to human-reviewed paths until measured usage separates the contenders [^8][^9].

Do not read output length as quality

A tempting shortcut is to assume the longer answer is the better one. The judges did not behave that way consistently. When discovering thematic topics, the answers judges picked were longer than the ones they passed over, 8,952 against 4,753 output tokens [^2].

When assigning sections to topic clusters the direction reversed, and the winning answers were shorter, 332 against 692 output tokens [^10][^11]. Verbosity cannot be used as a cross-task quality signal, so price on usable output rather than on generated tokens, and check length before treating any judged win as a quality lead.

The routing rule that falls out of this is simple to state: match the model to the exact topic step, not to the category. Suggestion, discovery, taxonomy generation and section assignment each have their own leader, and a leader on one has already been shown to fail on another.

Where to look next: the task-level pages behind these numbers are live, so check the current scores, intervals and sync-versus-batch costs for the exact step you plan to run at https://llm-bench.kapualabs.com/ and on the fronset.ai benchmark task pages linked in the sources below. Those numbers decide the route; this article only shows why a single best-model claim is the wrong question to ask.