Best Models for Content Summarization & Synthesis
GPT-5.4 Nano Beats GPT-5.5 at Structured Summaries, and Costs Less
If you send summarization work to the biggest model you can afford, this benchmark suggests you are paying more for worse output. On turning structured content into summaries, GPT-5.4 Nano scored higher than GPT-5.5 at a lower price per run, and several other cheap models land close behind it. The second finding is about mode rather than model: batch pricing takes roughly half off the bill, and batches come back in minutes rather than the advertised day.
How these numbers are produced
Every score is one judge grading one output against a published rubric on a 0–10 scale, and the ranked scores rest on 239,838 scored judgments collected between 2026-06-03 and 2026-08-21 on real production work [^1][^2]. A model counts as good enough on a task when it lands within 90% of the best measured model, with the bar raised to 95% or lowered to 80% on some tasks [^1][^2]. Confidence intervals matter when reading the table below: where two intervals overlap, treat the models as tied.
The small model wins the headline task
The table lists the structured-content summarization results the source quotes for the flagship, its nano sibling, and the cheap alternatives.
| Model | Quality (0–10) | Cost per task run |
|---|---|---|
| GPT-5.4 Nano | 9.29 (±0.09) [^3][^4] | $0.0019, batch [^3][^4] |
| GPT-5.6 Luna | 8.91 (±0.12) [^3][^4] | $0.0041, batch [^3][^4] |
| Moonshot Kimi K3 | 8.76 (±0.24) [^3][^4] | $0.11, sync [^3][^4] |
| MiniMax M3 | 8.42 (±0.16) [^3][^4] | $0.0014, sync [^3][^4] |
| GPT-5.5 | 6.83 (±0.31) [^3][^4] | $0.03, batch [^3][^4] |
The ordering is the surprise. The nano model is the top scorer of the group and the flagship is the lowest, at a higher per-run price than the nano model. Nothing about model size or family name predicted this; the task did.
Price status is worth a look before you copy these numbers into a budget. GPT-5.4 Nano’s cost is measured on this model’s own usage, while the GPT-5.6 Luna and Moonshot Kimi K3 costs are estimates with no qualifying usage history of their own [^3][^4]. MiniMax M3 is the cheapest model that still clears the 90% bar on this task, which makes it the natural pick when the request has to be interactive [^3][^4].
Cheap does not mean sloppy, but watch how it fails
Low price does not buy you a free pass on reliability, and the failure profile differs by model. On summarizing structured content, judged failure runs range from 22.1% for MiniMax M3 up to 52.1% for Gemini 3.5 Flash [^4]. Gemini 3.5 Flash fails mostly by invention and MiniMax M3 mostly by brevity [^4], and that distinction is operational: a too-short summary is obvious and can be retried, while an invented detail risks slipping through as an unnoticed wrong answer.
Answer rate is the other qualifier on any quality score. A summary score only applies when the model actually answers, and Claude Sonnet 5, for example, answers on 86% of attempts on structured summaries [^3][^4]. If your pipeline assumes every call returns something usable, a model with a strong mean can still leave a visible hole in your output.
Batch is half the price, and it is not overnight
The other lever is deferral. Running Claude Sonnet 5 on synthesizing multiple perspectives into a decision through batch costs $0.05 per task against $0.12 synchronously, 58% less for work that can wait [^5][^6]. The usual objection to batch is turnaround, and the data does not support it: half of all batches are back within 5.5 minutes and 75.1% within 15 minutes, against a published window of up to 24 hours [^1][^2].
That makes the routing rule simple. If the input is already structured and the summary can wait, run GPT-5.4 Nano in batch; if it must be interactive, run MiniMax M3 in sync. The one constraint is eligibility, since only 50 of 68 benchmark tasks are batch-eligible [^1][^2], so check the task page before you assume the discount applies.
Where to look next
These are two tasks out of a much larger set, and the cheapest qualifier changes with the mode you run and the quality bar you set. The live benchmark at https://llm-bench.kapualabs.com/ lets you set the bar you actually need and compare batch against sync pricing yourself, and the structured-content summarization task page at https://fronset.ai/benchmark/task/structured-content-summarization/ carries the per-model detail behind the table above. Re-run the comparison for your own task before locking a default.