Four tasks where the model you pick actually changes the result

Most tasks in our benchmark end in a near-tie between the top two models, which is why “route on cost” is usually the right call. But a handful of jobs break that pattern: community selection, metadata paragraph improvement and content-domain suggestion produce real gaps between the leaders, with specific models winning them head-to-head and other models failing them outright. This piece is about those jobs, and about how to tell whether one of yours belongs in that group.

The tasks that split the field

Every figure here comes from independent judgments of models answering the same live input. For each task the benchmark tracks how often the top two answers tie exactly, how often they finish within half a point, and the median gap between them on a 10-point scale. Those three numbers are the tell: when ties are rare and the median gap is wide, model choice is doing real work.

Community selection was the most differentiated task by tie rate, and the second content-domain setup has the widest median gap. In all four, the top two answers agree exactly less than one time in ten. These are the jobs where testing candidates on your own prompts is most likely to pay off.

What a tie looks like, for contrast

Identifying language is the opposite case: the top two answers scored exactly the same in 97.2% of 71 judged occasions, with a median gap of 0.0 [^8][^9]. Choosing a post for X tied exactly in 64.3% of 84 judged occasions and finished within half a point in 90.5% [^10][^11].

Extraction, matching and routine summarization behave the same way, with the top two routinely finishing within half a point of each other. In that band, swapping the best answer for the runner-up rarely changes judged quality, so price, latency and workflow fit should decide.

Who wins the differentiated jobs

Content-domain suggestion has clear leaders. Meta Muse Spark 1.1 was judged the best answer in 39 of the 45 occasions it competed in, 86.7%, every one of them against other models answering the same live input [^6][^7]. Thinking Machines Inkling Small was close behind, judged best in 28 of the 33 occasions it competed in, 84.8% [^7].

Selection work shows the same concentration. On selecting an appropriate subreddit, GPT-5.6 Sol was judged the best answer in 25 of the 29 occasions it competed in, 86.2% [^1][^2]. If those are your jobs, these are the models to start with rather than whichever model tops the general leaderboard.

The outliers to keep off those jobs

The flip side is that a model which leads one differentiated task can fail another. GPT-5.6 Sol, the subreddit leader above, scores 3.84 out of 10, ±0.1 after 30 independent judgments when suggesting a content domain [^12][^5].

Claude Sonnet 5 discovers thematic topics at 1.02 out of 10, ±0.08 after 30 independent judgments, which is not a near-miss but a failure mode sitting alongside otherwise 8 to 9-level performance [^13][^14]. Results like these are exactly what an aggregate leaderboard hides. Pairings like these belong off the critical path unless a reviewer or a different model is in the loop.

How to tell which of your tasks deserve a bake-off

Look at the tie rate and the median gap before you look at the absolute score. A high absolute score on identification or structuring work tells you several models can do the job; it does not tell you which one to deploy, because the head-to-head gap is too small for quality alone to justify paying more or accepting slower turnaround.

When the exact-tie rate drops under ten percent and the median gap climbs toward a full point, treat the task as model-sensitive. Run your own prompts through two or three candidates, start with the models that hold the highest best-answer share on the benchmark’s closest task, and check each candidate’s score on that specific task for a low outlier before committing. Everything else can go to the cheapest model that clears your quality bar.

Where to look next

The live benchmark at https://llm-bench.kapualabs.com/ lets you set the quality bar you actually need and compare batch against sync pricing for yourself, and each task page (the subreddit selection and content-domain suggestion pages in the sources below are good starting points) shows the tie rates and best-answer shares behind the numbers here. The method is documented at https://llm-bench.kapualabs.com/methodology/.