Best LLMs for Structured Content Summarization
Produces a comprehensive, source-grounded summary of supplied content, optionally organized around a caller-provided structure or analytical concerns. It can isolate the substantive material in noisy web-page text containing navigation, formatting remnants, advertisements, consen
Models
Frontier on this task: GPT-5.4 Nano at 9.24 / 10. Quality bar at 90%: 8.32.
point-estimate floor (CI low) · upper CI (less certain) · Bars sorted by blended cost; best-value model first. Greyed rows are MEDIUM+ models whose point estimate clears the bar but whose CI low does not.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| MiniMax M3 | 8.39 / 10 | 8.26 | $2.12 | best value |
| GPT-5.4 Nano | 9.24 / 10 | 9.15 | $2.30 | 1.1x more expensive |
| GLM-5.3 Flash | 8.60 / 10 | 8.42 | $2.42 | 1.1x more expensive |
| GPT-5.6 Luna | 8.89 / 10 | 8.78 | $2.73 | 1.3x more expensive |
| Tencent Hy3 | 8.67 / 10 | 8.50 | $3.74 | 1.8x more expensive |
| Thinking Machines Inkling Small | 8.35 / 10 | 8.04 | $6.86 | 3.2x more expensive |
| GPT-5.6 Terra | 8.34 / 10 | 8.16 | $12.80 | 6x more expensive |
| Meta Muse Spark 1.3 | 9.06 / 10 | 8.90 | $13.72 | 6.5x more expensive |
| Grok 4.6 | 8.44 / 10 | 8.11 | $20.19 | 9.5x more expensive |
| GLM-5.3 | 8.62 / 10 | 8.32 | $20.59 | 9.7x more expensive |
| GPT-5.6 Sol | 8.84 / 10 | 8.63 | $27.92 | 13x more expensive |
| Tencent Hy4 Preview | 8.77 / 10 | 8.45 | $40.92 | 19x more expensive |
| Qwen 3.8 Max | 8.36 / 10 | 7.99 | $54.13 | 25x more expensive |
| Moonshot Kimi K3 | 8.66 / 10 | 8.45 | $54.74 | 26x more expensive |
| Claude Opus 5 | 8.91 / 10 | 8.69 | $62.58 | 29x more expensive |
| Gemini 3.8 Flash | 7.95 / 10 | 7.46 | $5.57 | 2.6x more expensive |
| Qwen 3.8 Flash | 7.84 / 10 | 7.38 | $4.15 | 2x more expensive |
| Claude Haiku 4.5 | 7.19 / 10 | 7.02 | $7.65 | 3.6x more expensive |
| Claude Sonnet 5 | 8.10 / 10 | 7.98 | $15.79 | 7.4x more expensive |
| DeepSeek V4 Flash | 7.22 / 10 | 7.01 | $6.47 | 3.1x more expensive |
| DeepSeek V4 Pro | 7.33 / 10 | 7.12 | $25.84 | 12x more expensive |
| Gemini 3.1 Flash Lite | 6.05 / 10 | 5.83 | $0.37 | 82% cheaper |
| Gemini 3.5 Flash | 6.12 / 10 | 5.79 | $30.37 | 14x more expensive |
| Gemini 3.5 Flash Lite | 5.88 / 10 | 5.58 | $0.89 | 58% cheaper |
| NVIDIA Nemotron-3 Nano 30B-A3B | 5.79 / 10 | 5.33 | $1.01 | 52% cheaper |
| NVIDIA Nemotron-3 Ultra 550B | 7.36 / 10 | 7.11 | $4.69 | 2.2x more expensive |
| Qwen 3.7 Plus | 7.95 / 10 | 7.79 | $7.49 | 3.5x more expensive |
| Thinking Machines Inkling | 8.30 / 10 | 8.02 | $27.58 | 13x more expensive |
| NVIDIA Nemotron-3 Super 120B | 7.08 / 10 | 6.78 | $2.73 | 1.3x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| MiniMax M3 ★ OpenRouter | 8.39 / 10 CI [8.26, 8.52] | RANKED | $2.12 | best value | batch |
| GPT-5.4 Nano best OpenAI | 9.24 / 10 CI [9.15, 9.33] | RANKED | $2.30 | 1.1x | batch |
| GLM-5.3 Flash Z.AI | 8.60 / 10 CI [8.42, 8.78] | RANKED | $2.42 | 1.1x | batch |
| GPT-5.6 Luna OpenAI | 8.89 / 10 CI [8.78, 9.00] | RANKED | $2.73 | 1.3x | batch |
| Tencent Hy3 OpenRouter | 8.67 / 10 CI [8.50, 8.84] | RANKED | $3.74 | 1.8x | batch |
| Thinking Machines Inkling Small OpenRouter | 8.35 / 10 CI [8.04, 8.65] | MEDIUM | $6.86 | 3.2x | batch |
| GPT-5.6 Terra OpenAI | 8.34 / 10 CI [8.16, 8.52] | RANKED | $12.80 | 6x | batch |
| Meta Muse Spark 1.3 OpenRouter | 9.06 / 10 CI [8.90, 9.21] | RANKED | $13.72 | 6.5x | batch |
| Grok 4.6 xAI | 8.44 / 10 CI [8.11, 8.78] | MEDIUM | $20.19 | 9.5x | batch |
| GLM-5.3 Z.AI | 8.62 / 10 CI [8.32, 8.91] | HIGH | $20.59 | 9.7x | batch |
| GPT-5.6 Sol OpenAI | 8.84 / 10 CI [8.63, 9.05] | HIGH | $27.92 | 13x | batch |
| Tencent Hy4 Preview OpenRouter | 8.77 / 10 CI [8.45, 9.09] | MEDIUM | $40.92 | 19x | batch |
| Qwen 3.8 Max Alibaba Cloud (DashScope) | 8.36 / 10 CI [7.99, 8.72] | MEDIUM | $54.13 | 25x | batch |
| Moonshot Kimi K3 Moonshot AI | 8.66 / 10 CI [8.45, 8.88] | HIGH | $54.74 | 26x | batch |
| Claude Opus 5 Anthropic | 8.91 / 10 CI [8.69, 9.13] | HIGH | $62.58 | 29x | batch |
Overpay shows how much more you pay than the best-value model that clears the quality bar (marked ★) — the best-value good-enough option. "16x" means you overpay 16× — 16× that reference for no quality benefit above the bar. Typical call shape for this task: 3067 input tokens → 2800 output tokens, EMA-tracked from production traffic. Cost is the observed, all-in $ per 1,000 task runs: each model's own measured usage on this task — output verbosity, thinking/reasoning tokens, cache reads and writes, and the spend on its billed failures — priced at current list rates and adjusted by the billing overhead we actually reconcile against provider invoices. Models that answer tersely cost what they actually cost; models that think at length pay for it. Not comparable to providers' advertised $/1M list rates — this is what running the task costs, not a per-token price.
Evaluation rubric
Judge correct isolation of primary content, removal of irrelevant page furniture and promotional noise, source fidelity, coverage of material facts and arguments, preservation of quantitative details, corrections, disclosures, warnings, and caveats, usefulness of organization, absence of invention, and appropriate compression. Penalize summaries that repeat navigation or advertisements, omit qualifying material as boilerplate, or reward brevity at the expense of decision-relevant information. Schema validity is deterministic.
Prompt templates
This is a pooled capability — 3 prompt families share it. The pair shown first is the most frequently used in production.
LLMB_STRUCTURED_CONTENT_SUMMARIZATION_SYSTEM +
LLMB_STRUCTURED_CONTENT_SUMMARIZATION_USER
(31339 calls in window)
System prompt
Identify and summarize the supplied content’s substantive material for downstream analysis. Treat layout and formatting as possible structural evidence, not automatically as content. Exclude page chrome, navigation, cookie or consent banners, subscription prompts, advertisements, promotional inserts, unrelated recommendations, social widgets, duplicated headers or footers, formatting artifacts, and other material unrelated to the primary content. Preserve bylines, dates, captions, footnotes, corrections, disclosures, warnings, disclaimers, and source context when they materially qualify or explain the substantive content. Prioritize faithful coverage of material facts, arguments, evidence, caveats, and conclusions over brevity. Use summary_structure to organize and prioritize content when supplied, but do not invent coverage for empty areas. Preserve important names, dates, numbers, units, uncertainty, and disagreements, and distinguish the source’s assertions from the summarizer’s organization. Treat empty optional values as absent and return only the requested result. Your response must conform exactly to this output schema: {schema_json_string}.
User prompt
Inputs — summary_structure: {summary_structure}; content_title: {content_title}; content_url: {content_url}; content_type: {content_type}; content_text: {content_text}. Use only these inputs to complete the task defined by the system prompt.
CONTENT_SUMMARIZATION_SYSTEM_PROMPT +
CONTENT_SUMMARIZATION_USER_PROMPT
(451 calls in window)
System prompt
You are an expert content analyst. Your task is to create a comprehensive summary of the provided content, extracting ALL information that is relevant to the report structure requirements provided in the user message.
## Instructions:
1. Extract ALL facts, data points, quotes, and insights relevant to ANY of the report chapters listed in the user message
2. Do NOT impose arbitrary length limits - capture everything relevant
3. Focus on substance over style - preserve key details, statistics, and specific claims
4. Remove irrelevant tangents, formatting artifacts, and social media noise
5. Write in clear, objective prose
6. If the content is highly relevant, your summary may be very long - that's expected
Your summary will be used for topic clustering and synthesis, so completeness is more important than brevity.
## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```User prompt
Please summarize the following content, extracting all information relevant to the report structure requirements.
## Report Structure Requirements:
{report_structure}
## Content Metadata:
- Title: {content_title}
- Source: {content_url}
- Type: {content_type}
## Content:
{content_text}
## Required Output Format:
The required JSON output schema is provided in the system prompt.
JSON_REPAIR_SYSTEM +
JSON_REPAIR_USER
(8 calls in window)
System prompt
You are a JSON repair tool. The user gives you malformed or partial model output and a JSON Schema. Return ONLY a single valid JSON object that satisfies the schema, salvaging as much real content from the input as possible. Do not invent data for fields the input doesn't support — use the schema's allowed empty/null values. Output the JSON object only: no prose, no markdown, no code fences.
User prompt
JSON Schema:
{schema_json}
Malformed output to repair:
{raw_text}
Return only the corrected JSON object.