Best Models for Content Summarization & Synthesis
Content Summarization & Synthesis: Quality Depends on the Job
For this category, “good enough” is not a single headline score. A production model must deliver sufficiently strong output for the intended workflow—whether turning source material into analysis, summarizing content, generating sections, or filing chunks—while also meeting required standards for clean output, completed pipelines, cost, and latency. The supplied task-level extracts do not specify how many tasks this category contains or how many models are measured on them, so those totals cannot be reported without unsupported inference. The measurement is continuous and ongoing; the figures cited below were published on a single date and represent a standing programme, not a momentary run.
Where the results are strongest
The evidence splits clearly by task type. For straightforward content summarization, gpt-5.4-nano is the highest reported result at 9.33 out of 10 at $0.0045 per synchronous task run [^3]. That exceeds gpt-5.6-sol at 8.95 out of 10 and $0.04 in batch mode [^3], gpt-5.6-luna at 8.92 out of 10 and $0.0048 in synchronous mode [^3], kimi-k3 at 8.74 out of 10 at $0.05 [^3], meta/muse-spark-1.1 at 8.39 out of 10 at $0.02 [^3], and grok-4.5 at 8.42 out of 10 [^3]. Because these rows mix synchronous and batch execution, they support gpt-5.4-nano for the measured summarization workload rather than a universal cost-adjusted ranking.
For broader synthesis and analysis, kimi-k3 has the highest reported score at 9.09 out of 10 and $0.18 per task run [^1][^6]. Alternatives below it include grok-4.5 at 8.84 out of 10 [^1], Claude Opus 4-8 at 8.73 out of 10 [^1], Claude Sonnet 5 at 8.72 out of 10 at $0.13 in synchronous mode [^1], meta/muse-spark-1.1 at 8.65 out of 10 at $0.06 [^1], gpt-5.6-luna at 8.20 out of 10 at $0.0065 [^1], and gpt-5.6-sol at 8.17 out of 10 at $0.08 in batch mode [^1]. kimi-k3’s performance across related work supports its use when summarization must lead into longer-form content: 8.74 out of 10, ±0.27 for summarization [^3], 9.34 out of 10, ±0.08 for generating sections [^2][^5], and 9.09 out of 10, ±0.23 for synthesis [^1][^6]. Additional reported options include Tencent’s hy3 at 8.70 out of 10 for summarization [^3], thinkingmachines/inkling-small at 8.8 out of 10 for synthesizing analysis [^1], and Qwen3.7-plus at 8.64 out of 10, ±0.15 for synthesis analysis and 7.97 out of 10, ±0.17 for content summarization [^1][^6][^3].
Where results are mixed
Category-level verdicts disagree at the task level, and the current evidence changes the practical recommendations from the prior cycle. The previous analysis emphasized Claude Opus for demanding synthesis and Inkling-small for summarization; this cycle reaffirms that task-level disagreement dominates, but it elevates kimi-k3 to the cited synthesis lead and gpt-5.4-nano to the cited summarization lead. Within the Claude family specifically, claude-opus-4-8 scores 8.33 out of 10, ±0.18 at $0.06 per synchronous task run [^3], while claude-haiku-4-5 scores 7.27 out of 10, ±0.16 at $0.02 per synchronous task run [^3], and claude-sonnet-5 scores 8.05 out of 10, ±0.23 for summarization with an 89% answer rate [^3]. That makes Opus the cited high-quality family option, Sonnet the completion-reliable alternative at a lower quoted score on this specific task, and Haiku a tighter-budget option with a materially lower measured result.
The same divergence appears across execution modes for the same low-cost provider: gpt-5.4-nano achieves 8.41 out of 10 at $0.0042 in batch mode on chunk-level filing analysis [^7], but only 6.02 out of 10 at $0.002 in batch mode on broader SEC filing analysis [^4]. A different quality bar, synchronous versus batch execution, tolerance for unanswered attempts, or a latency requirement can change the choice entirely.
Where it is weak or not competitive
Measured but not competitive for the main use cases include:
- claude-haiku-4-5 at 7.27 out of 10 for summarization [^3], well below the family’s Opus option at 8.33 [^3].
- deepseek-v4-flash at 7.1 out of 10 for summarization [^3], despite stronger results in structured extraction and document analysis.
- Qwen3.7-plus at 7.97 out of 10 for content summarization [^3], below gpt-5.4-nano and kimi-k3.
- gpt-5.6-sol at 8.17 out of 10 for synthesis [^1] and gpt-5.6-luna at 8.20 out of 10 [^1], both below kimi-k3, Claude Opus, and Claude Sonnet for synthesis, so they should be treated as value fallbacks rather than leaders.
For synthesis specifically, the gap is meaningful: kimi-k3 at 9.09 versus gpt-5.6-luna at 8.20 is nearly a full point, and the provider’s quoted cost differs by roughly an order of magnitude. The negative or inconclusive results are publishable: a model can be weakest in this category yet still serve narrower requirements, which is why task-level checks matter.
Head-to-head alternatives
A structured comparison of the cited leaders shows where quality and cost diverge.
| Model (display name) | Work / Task cited | Score (quoted) | Cost / Mode (quoted) | Evidence |
|---|---|---|---|---|
| gpt-5.4-nano | Content summarization | 9.33 / 10 [^3] | $0.0045 sync [^3] | Highest reported summarization |
| kimi-k3 | Content summarization | 8.74 ±0.27 / 10 [^3] | $0.05 [^3] | Stronger than Opus/Haiku here |
| claude-opus-4-8 | Content summarization | 8.33 ±0.18 / 10 [^3] | $0.06 sync [^3] | High-end family option |
| claude-sonnet-5 | Content summarization | 8.05 ±0.23 / 10 [^3] | Not quoted for cost | 89% answer rate [^3] |
| kimi-k3 | Synthesis / analysis | 9.09 ±0.23 / 10 [^1][^6] | $0.18 [^1][^6] | Highest reported synthesis |
| grok-4.5 | Synthesis / analysis | 8.84 / 10 [^1] | Not quoted | Second-tier synthesis |
| claude-opus-4-8 | Synthesis / analysis | 8.73 / 10 [^1] | Not quoted | Near-top synthesis |
| claude-sonnet-5 | Synthesis / analysis | 8.72 / 10 [^1] | $0.13 sync [^1] | Balanced quality/cost |
| gpt-5.6-luna | Synthesis / analysis | 8.20 / 10 [^1] | $0.0065 [^1] | Lowest-cost above ~8.0 |
In content summarization, the quality gap between the top score (gpt-5.4-nano at 9.33) and the relevant high-end comparison (claude-opus-4-8 at 8.33) is about one point, while the cost gap is roughly 13× ($0.0045 versus $0.06) [^3]. Within synthesis, the gap between kimi-k3 (9.09 at $0.18) and gpt-5.6-luna (8.20 at $0.0065) is nearly 0.9 points for a cost multiple near 28× [^1][^6][^1]. For literal summarization, thinkingmachines/inkling-small at 8.8 for synthesizing analysis [^1] and 8.57 in prior-cycle summarization (not restated as current source) support using it for synthesis-to-analysis workflows rather than as a universal replacement for either gpt-5.4-nano or kimi-k3.
Reliability caveats: what the evidence can and cannot settle
Only what the source material states. Confidence intervals are reported for kimi-k3 (±0.27 summarization [^3], ±0.08 sections [^2][^5], ±0.23 synthesis [^1][^6]), for Claude Opus (±0.18 [^3]), for Claude Haiku (±0.16 [^3]), for Claude Sonnet summarization (±0.23 [^3]), and for Qwen3.7-plus (±0.15 synthesis [^1][^6], ±0.17 summarization [^3]). The evidence does not specify sample size or measurement volume for each interval, so statistical separation between every closely spaced pair cannot be assumed. The benchmark is continuous; these are publication-date figures, not a one-off measurement. Because some comparisons use synchronous execution and others batch, cost and latency comparisons across rows are approximate rather than strictly comparable. Clean-output rates, salvage rates, and end-to-end success rates are not available for every cited model, so production usability must be validated against the exact pipeline rather than inferred from category-level quality alone.
Routing conclusion and rule
The best quality on quality alone depends on the task: gpt-5.4-nano for summarization, kimi-k3 for synthesis. The cheapest model that still clears a practical bar around 8.0 is gpt-5.4-nano itself for summarization, and gpt-5.6-luna at $0.0065 for synthesis at 8.20 [^1]. For document-heavy work, do not choose from category verdict alone: validate chunk-level filing versus broad SEC filing separately. The routing rule below can be applied without reading the rest of this article: Use gpt-5.4-nano for straightforward content summarization when synchronous cost and unanswered-attempt handling fit; use kimi-k3 for synthesis that extends into generated sections if $0.18 per run is justified; fall back to gpt-5.6-luna for high-volume synthesis when ~8.2 quality is sufficient; prefer claude-opus-4-8 over claude-haiku-4-5 when summarization must exceed ~8.3; use claude-sonnet-5 when an 89% answer rate is controlling; and test document-heavy variants separately because the same model can diverge sharply between chunk-level and broad filing tasks.
You can set the quality bar your workflow actually needs and compare batch against sync pricing yourself at https://llm-bench.kapualabs.com/, where the continuous measurement lets you verify whether gpt-5.4-nano, kimi-k3, or gpt-5.6-luna fits your production pipeline. Check the numbers there before turning this rule into policy—quality depends on the job, and the live benchmark lets you measure it precisely.