Best LLMs for Reference Preserving Analytical Writing
Produces analytical prose from supplied claims and source material while preserving references according to a caller-supplied reference protocol. Illustrative uses include producing a cited vendor due-diligence report, contract analysis, technical assessment, market-intelligence
Run this task on Fronset — request an invitation
Models
Frontier on this task: Grok 4.7 at 9.37 / 10. Quality bar at 90%: 8.43.
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 does not clear the bar.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| Xiaomi MiMo V2.6 Pro | 8.52 / 10 | 8.12 | $9.46 | best value |
| GLM-5.3 Flash | 9.24 / 10 | 9.03 | $39.73 | 4.2x more expensive |
| Claude Sonnet 5.5 | 8.67 / 10 | 8.35 | $53.46 | 5.7x more expensive |
| Grok 4.7 | 9.37 / 10 | 9.14 | $118.87 | 13x more expensive |
| Claude Opus 5.5 | 8.66 / 10 | 8.29 | $120.03 | 13x more expensive |
| Moonshot Kimi K3 | 9.08 / 10 | 8.95 | $178.42 | 19x more expensive |
| GLM-5.3 | 8.88 / 10 | 8.39 | $509.24 | 54x more expensive |
| Qwen 3.8 Max | 8.20 / 10 | 7.80 | $223.27 | 24x more expensive |
| GPT-6 Luna | 8.04 / 10 | 7.71 | $2.92 | 69% cheaper |
| Gemini 3.5 Flash Lite | 6.96 / 10 | 6.55 | $6.58 | 30% cheaper |
| NVIDIA Nemotron-3 Nano 30B-A3B | 5.02 / 10 | 4.65 | $4.94 | 48% cheaper |
| GPT-6 Sol | 8.28 / 10 | 7.98 | $28.66 | 3x more expensive |
| DeepSeek V4.1 Flash | 8.12 / 10 | 7.71 | $6.08 | 36% cheaper |
| Meta Muse Spark 1.3 | 7.96 / 10 | 7.53 | $64.17 | 6.8x more expensive |
| Gemini 3.8 Flash | 7.95 / 10 | 7.50 | $20.29 | 2.1x more expensive |
| Thinking Machines Inkling Small | 7.89 / 10 | 7.47 | $21.89 | 2.3x more expensive |
| Xiaomi MiMo V2.6 Flash | 8.18 / 10 | 7.77 | $2.17 | 77% cheaper |
| Claude Haiku 5.5 | 8.35 / 10 | 8.00 | $1.30 | 86% cheaper |
| Qwen 3.7 Plus | 8.12 / 10 | 7.85 | $25.10 | 2.7x more expensive |
| Thinking Machines Inkling | 8.42 / 10 | 8.07 | $73.01 | 7.7x more expensive |
| Tencent Hy3 | 7.28 / 10 | 6.84 | $6.30 | 33% cheaper |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| Xiaomi MiMo V2.6 Pro ★ Xiaomi MiMo | 8.52 / 10 CI [8.12, 8.93] | MEDIUM | $9.46 | best value | batch |
| GLM-5.3 Flash Z.AI | 9.24 / 10 CI [9.03, 9.44] | HIGH | $39.73 | 4.2x | batch |
| Claude Sonnet 5.5 Anthropic | 8.67 / 10 CI [8.35, 9.00] | MEDIUM | $53.46 | 5.7x | batch |
| Grok 4.7 best xAI | 9.37 / 10 CI [9.14, 9.60] | HIGH | $118.87 | 13x | batch |
| Claude Opus 5.5 Anthropic | 8.66 / 10 CI [8.29, 9.04] | MEDIUM | $120.03 | 13x | batch |
| Moonshot Kimi K3 Moonshot AI | 9.08 / 10 CI [8.95, 9.21] | RANKED | $178.42 | 19x | batch |
| GLM-5.3 Z.AI | 8.88 / 10 CI [8.39, 9.37] | MEDIUM | $509.24 | 54x | 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: 30256 input tokens → 9945 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 factual grounding, correct and complete reference placement, reference identity preservation, analytical coherence, synthesis rather than concatenation, requirement coverage, and clarity. Any invented or misattached reference is a severe defect.
Output schema
Every answer on this task is checked against this JSON Schema, whichever model wrote it. An answer that doesn't fit counts as a model failure, and the call is retried on another model.
{
"properties": {
"report_markdown": {
"description": "The full synthesized report formatted in Markdown.",
"title": "Report Markdown",
"type": "string"
}
},
"required": [
"report_markdown"
],
"title": "SynthesizedOutputMarkdownOnly",
"type": "object"
}Prompt templates
The published template for this task.
default ·
LLMB_REFERENCE_PRESERVING_ANALYTICAL_WRITING_SYSTEM +
LLMB_REFERENCE_PRESERVING_ANALYTICAL_WRITING_USER
(8283 calls in window)
System prompt
Synthesize the supplied evidence into clear analysis. Apply reference_profile exactly: preserve reference identifiers, attach them to supported assertions, and perform no create, renumber, merge, or reuse operation unless the profile explicitly permits it. Separate source-backed facts from interpretation and omit unsupported conclusions. Use only the citation syntax specified by reference_profile; if none is specified, do not invent one. 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 — author_voice_section: {author_voice_section}; requirement: {requirement}; sources: {sources}; claims_text: {claims_text}; subject_name: {subject_name}; subject_code: {subject_code}; chapter_title: {chapter_title}; chapter_requirement: {chapter_requirement}; category: {category}; claim_count: {claim_count}; topic_name: {topic_name}; topic_description: {topic_description}; synthesis_text: {synthesis_text}; reference_profile: {reference_profile}. Use only these inputs to complete the task defined by the system prompt.