Best Models for Long-form Content Generation
Most top results in the benchmark’s long-form writing tasks are statistical ties, and the top score changes from one task to the next. Inside a tie, a higher score is not a measurable gain, so price and how often the model answers should decide. On reference-preserving analytical writing, GLM-5.3 Flash ties Kimi K3 at a lower price but answers on only 52% of attempts [^1][^2].
What the benchmark measures, and what counts as a tie
None of the results here scores long-form output as one task. They come from nearby writing jobs instead, such as report outlines, executive summaries, publication titles, newsletter copy and reference-preserving analytical writing, where the text keeps its reference markers intact. None of them reports output length, coherence across a full document or accuracy over a long text.
Each quality score is out of 10 and comes with a ± margin of uncertainty, its confidence interval. When two models’ intervals overlap, the benchmark cannot separate them, and the result is a tie, whatever the point scores say.
A cheaper model can tie the top score
Kimi K3 posts the highest quoted score on report outlines, 9.65 out of 10, ±0.09 [^3][^4][^5][^6]. Grok 4.5 follows at 9.52, ±0.04 [^3][^4][^5], and Thinking Machines Inkling Small at 9.46, ±0.11 [^3][^6]. With intervals this close, the three read as a tie, and Kimi K3 is higher only on the point score.
Price is where Grok 4.5 and Kimi K3 differ. Grok 4.5 is the cheapest outline result that clears 9 out of 10, at $0.02 per run [^3][^4][^5], while Kimi K3 costs $0.10 per run on its own usage [^3][^4][^5][^6][^7]. These per-run costs are not on a common price basis, so treat them as quoted figures rather than a like-for-like comparison.
A tied score can hide a low answer rate
On reference-preserving analytical writing, GLM-5.3 Flash scores 9.22 out of 10, ±0.2 [^2], and Kimi K3 scores 9.1, ±0.18 [^1][^2]. Their intervals overlap, so they are tied. GLM-5.3 Flash costs $0.04 per run [^2], against an estimated $0.33 for Kimi K3 [^1][^2].
The catch is the answer rate: the share of attempts on which a model returns an answer at all. GLM-5.3 Flash answers on only 52% of attempts on this task, and its score describes only the calls that returned [^2]. Whether that is acceptable depends on the job.
Publication title packages show the same trap more starkly. GPT-6 Sol scores 8.98 and GPT-6 Luna scores 8.61, but both answer on only 10% of attempts, so those scores describe a small minority of calls [^8].
Kimi K3’s answer rate on reference-preserving analytical writing is not given here, so the pricier option is not a proven fallback either. A model’s answer rate on one task may not carry over to another: Qwen 3.8 Max answers on 65% of executive summary attempts [^5] and 74% of structured summary attempts [^9].
What to do with this
- Shortlist by task: keep every model whose interval overlaps the highest scorer’s on the exact job you need.
- Compare answer rates on that same task before you compare prices.
- If a path depends on every call returning, give it a fallback model.
- Test on full-length documents before relying on these per-task scores for a whole report.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar your deliverable needs. Check these figures there before committing a model to a production path.