Models Tie on Language ID but Split on Content Domains

On identifying what language a text is written in, the leading models are tied at or near 10 out of 10 [^1]. On suggesting which content domain a piece belongs to, the same field spreads out, and only a small group pulls clearly ahead [^2]. On the first kind of task, choosing a model is a question of price and latency; on the second, it is a question of quality.

Language identification is a tie at the top

Three models sit within each other’s intervals at the top: Claude Sonnet 5.5 at 10.0 (±0.11 after 46 independent judgments), DeepSeek V4.1 Flash at 9.97 (±0.07 after 100) and Moonshot Kimi K3 at 9.95 (±0.05) [^1]. Because the intervals overlap, that is a tie, not a ranking. NVIDIA Nemotron-3 Nano 30B-A3B is close behind at 9.82 after 459 judgments [^1].

The head-to-head judging, where models answer the same live input, says the same thing. In one set of 62 judged occasions, the top two answers scored exactly the same every time, with a median gap of 0.0 on a 10-point scale [^1]. A separately tracked language-detection task found identical top-two scores in 97.2% of 71 occasions [^3].

Language identification is not the only task like this. When translating batches of text, the top two finished within half a point in 91.4% of 70 judged occasions, with a median gap of 0.1 [^4]. When choosing a portfolio of social posts, they scored identically in 70.0% of 60 [^5]. On work like this, the score cannot tell you which model to pick.

Content domain suggestion splits the field

Suggesting a content domain is the weakest result for nearly every model that reports it. Most models measured land between 3.49 and 5.27, including Claude Sonnet 5.5 at 4.73, the same model that ties for first on language identification [^2]. Where intervals are quoted they are narrow, so these are measured weaknesses, not noise.

Three results sit well above that cluster: Thinking Machines Inkling Small at 7.37, Claude Opus 5.5 at 6.92 and Claude Haiku 5.5 at 6.26 [^2]. Model family is a poor guide here. Thinking Machines Inkling, without the Small, scores 4.1, and Claude Haiku 5.5 outscores Claude Sonnet 5.5 [^2].

The head-to-head judging confirms the spread. In one measurement, the top two answers were identical in only 8.0% of 87 judged occasions, with a median gap of 1.9 points [^6]. A second found 12.7% of 63 identical and a median gap of 0.8 [^2]. The two disagree on how wide the gap is, but both show far more separation than language identification.

Direct comparisons point the same way. Thinking Machines Inkling Small was judged the best answer in 28 of the 33 occasions it competed in, and meta/muse-spark-1.1 in 39 of 45, each against other models on the same live input [^6]. These are win shares, not scores. meta/muse-spark-1.1 is a different variant from Meta Muse Spark 1.3, which scores 3.92 [^2], so the two results do not contradict each other.

What to do with it

For language identification, batch translation and social-post selection, treat the leading models as interchangeable on quality and decide on price and latency.
For content domain suggestion, start with Thinking Machines Inkling Small or meta/muse-spark-1.1 [^2][^6]. Most other models sit in the low cluster, so here the model choice is the quality decision.
For any other task, check how often the top two answers tie before spending effort comparing models. If they usually tie, the comparison is about cost.

The caveat that matters is drift. Since 2026-10-16, both Inkling variants lost 0.22 points, Claude Opus 5.5 lost 0.28 and Claude Haiku 5.5 lost 0.29 on content domain suggestion [^2]. The three leaders still sit above the cluster, but the margin is moving, so re-check it before routing traffic on it.

The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar you need and compare batch against sync pricing yourself. Start by checking whether your task behaves like language identification, where the leading models are interchangeable, or like content domain suggestion, where only a few do. The method is documented at https://fronset.ai/benchmark/methodology/.