A Three-Way Tie for Best Long-Form Writer. Let Price Decide.
A Three-Way Tie for Best Long-Form Writer. Let Price Decide.
On the Fronset benchmark’s long-form writing work, the top of the table is a tie, not a leader. On reference-preserving analytical writing, the task closest to sustained, source-grounded drafting, three models sit inside overlapping score ranges, and so does a broad band below them [^1]. When the scores cannot separate models, price and pricing mode are what is left to decide on.
Three leaders, one tie
GLM-5.3 Flash scores 9.22 (±0.2), Grok 4.7 scores 9.19 (±0.21) and Moonshot Kimi K3 scores 9.1 (±0.13), all out of 10 [^1]. The ± figure is the range reported around each score, and when two ranges overlap, the difference between them cannot be treated as real. That is the case here: the 0.03-point gap between GLM-5.3 Flash and Grok 4.7 sits inside the overlap, and the 0.12-point gap to Kimi K3 is also within overlapping ranges [^1]. Calling any one of them “best on quality” would imply a ranking the numbers do not support.
Direct comparisons point the same way. When two answers were compared on this task, the top two finished within half a point in 88.4% of 69 judged occasions, with a median gap of 0.1 and exact ties in 24.6% [^1]. The models in those comparisons are not named, so this shows how close the finishes are rather than who won them.
The middle band is tied too
Below the leaders, a broad middle band runs from Claude Opus 5.5 at 8.64 (±0.39) down to Qwen 3.7 Plus at 8.12 (±0.27) [^1]. The ranges overlap across most of it, so these models are tied on reference-preserving writing. Claude Haiku 5.5 sits inside that band at 8.37 (±0.37) [^1].
Report executive summaries, another long-form job, show the same shape at a lower level. The top quoted score is GLM-5.3 at 8.6 (±0.18) at $0.29, and GLM-5.3 Flash scores 8.34 (±0.31) at $0.03; the benchmark reads the pair as a tie [^2].
When scores tie, price decides
At the top of reference-preserving writing, GLM-5.3 Flash costs $0.04 per task run against $0.12 for Grok 4.7, and no cost is reported for Kimi K3 [^1]. One caution sits outside the score: GLM-5.3 Flash returns an answer on only 52% of attempts [^1], so a path that needs an answer on every call needs a fallback.
In the middle band, too, cost is what separates the tied models. Claude Haiku 5.5 costs $0.0046 sync and $0.0024 in batch mode, against $0.29 sync for Claude Opus 5.5 [^1]. Both sit in the tied band, so on this task the difference between them is price, not a measured quality gap.
Batch mode, for work that can wait, is a second lever. For the Claude models quoted on this work, the score is the same in batch and sync, so mode changes price and not measured quality [^1]. The stated batch savings are 68% for Claude Opus 5.5 and 75% for GPT-6 Luna [^1].
What to do about it
Set the quality bar your deliverable needs, treat models whose ranges overlap as tied, and take the cheapest model that clears the bar rather than the one with the highest point score. Where the work can wait, run it in batch mode. Confirm the batch score for your own model, since the same-score finding is stated only for the Claude models quoted here.
Test on a full-length document before committing. No result in this category is labelled by output length, and the evidence does not establish quality across a full long document, so no result here should be carried over to a full-length report without testing.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your deliverable actually needs and compare batch against sync pricing yourself. The measurement is continuous, so check these figures there before committing a model to a production path.