What Changed This Snapshot?
Much of what changed in this benchmark snapshot is narrower confidence intervals, which mean a score is known more precisely, not that a model got better. Most score moves are small declines that vary by task, and none of them establishes a winner. If you pick or swap models when a benchmark updates, these are easy changes to misread.
Moving up a confidence level is not an upgrade
A confidence interval is the ± figure after a score. GLM-5.3 Flash’s interval on structured content summarisation, for instance, has narrowed from ±0.25 to ±0.18 since 2026-09-05 [^1].
Confidence levels, such as medium, high and ranked, describe how settled a ranking is. Meta Muse Spark 1.3 moved up a level on three tasks: from medium to high on two [^2][^3], and from high into ranked confidence on the third [^4]. In each case the interval closed enough to justify the move, but none of the three tells you the model scored better.
A score and a confidence level can even move in opposite directions. On catalyst and scenario analysis, Qwen 3.8 Max’s score rose in this snapshot, yet it was demoted from high to medium confidence on that same task [^5].
Scores moved task by task
Tencent Hy4 Preview had the largest single movement in the snapshot: down 0.71 points on public engagement reply review, to 8.4 on a 10-point scale, since 2026-09-24 [^6]. Whether that reflects the model or the measurement is not established.
Yet in the same snapshot, Tencent Hy4 Preview made small gains on document work: 0.08 points on profiled document section analysis, to 8.98 [^7], and 0.08 points on authoritative source selection, to 8.13 [^8].
Some declines were shared across models on a single task. On bench hypothesis post generation, scores fell for MiniMax M3 [^4][^9], Nemotron-3 Nano 30B-A3B [^9], Thinking Machines Inkling [^4][^9], Tencent Hy3 [^9] and Moonshot Kimi K3 [^4][^9]. For a team with a high quality bar, that is the clearest place to retest before relying on any of these models.
A move inside the interval is not yet a change
Meta Muse Spark 1.3 gained 0.07 points on short post batch relevance scoring, to 6.95 [^10], and lost 0.15 points on bench hypothesis post generation, to 8.68 [^4][^9]. Neither movement is shown to exceed normal variation.
The same goes for gaps between models. On structured content summarisation, Tencent Hy4 Preview lost 0.14 points, to 8.63 [^1], and now sits level with Tencent Hy3’s 8.7 [^1][^11]. Their intervals, ±0.27 and ±0.18, overlap [^1][^11], so the two are tied on that task.
How to read the next update
- Don’t upgrade or drop a model just because its confidence level moved.
- Count a score change only once it clears the interval, and read overlapping intervals as a tie.
- Re-measure the task you run before committing to a model for it, especially where a score fell sharply.
Check the live benchmark at https://fronset.ai/benchmark/ to set the quality bar your workload needs.