Cost mode:

Category: Infrastructure & Utility · Rail: absolute · Typical I/O: 586→1985 tokens

Models

Frontier on this task: Claude Sonnet 5 at 9.42 / 10. Quality bar at 90%: 8.48.

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.

ModelQuality scoreCI lowCost / 1k runsvs best value
Claude Sonnet 59.42 / 109.25$1.80best value
Gemini 3.5 Flash9.10 / 108.90$6.333.5x more expensive
NVIDIA Nemotron-3 Ultra 550B8.48 / 108.19$2.011.1x more expensive
DeepSeek V4 Pro8.15 / 107.86$1.991.1x more expensive
Gemini 3.1 Flash Lite7.53 / 107.29$0.2984% cheaper
GPT-5.4 Nano7.10 / 106.77$0.2387% cheaper
GPT-5.6 Luna6.56 / 106.34$0.1492% cheaper
Qwen 3.7 Plus7.41 / 107.20$5.593.1x more expensive
Tencent Hy36.83 / 106.62$0.1890% cheaper
GPT-5.6 Terra6.56 / 106.36$1.0741% cheaper
Thinking Machines Inkling Small6.74 / 106.39$1.821x more expensive
Thinking Machines Inkling6.46 / 106.24$4.552.5x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
Claude Sonnet 5 best Anthropic9.42 / 10 CI [9.25, 9.58]RANKED$1.80best valuebatch
Gemini 3.5 Flash Gemini9.10 / 10 CI [8.90, 9.30]HIGH$6.333.5xbatch

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: 586 input tokens → 1985 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 preservation, clarity, fluency, concision, appropriate metadata tone, and absence of invented methodology or stronger certainty. Exact protected-value preservation should be deterministic where possible.

Prompt templates

This is a pooled capability — 2 prompt families share it. The pair shown first is the most frequently used in production.

METADATA_PARAGRAPH_IMPROVEMENT_SYSTEM_PROMPT + METADATA_PARAGRAPH_IMPROVEMENT_USER_PROMPT (275 calls in window)

System prompt

You are an expert editor specializing in creating polished, professional metadata paragraphs for research reports.

Your task is to take a draft metadata paragraph that describes the research period and sources, and improve its wording to make it:
- More professional and polished
- Clear and concise
- Reader-friendly
- Accurate to the data provided

IMPORTANT GUIDELINES:
1. Preserve all factual information (dates, numbers, source types)
2. Do NOT add information that wasn't in the original
3. Do NOT change the meaning or facts
4. Keep the paragraph approximately the same length
5. Maintain a professional, informative tone
6. Start with a clear statement about the research period
7. Include source count and types naturally in the flow

The improved paragraph should be suitable for inclusion at the beginning of a professional research report or analysis document.

User prompt

Please improve the wording of this metadata paragraph while preserving all factual information:

{draft_paragraph}

Key facts to preserve:
- Date range: {date_from} to {date_to}
- Number of sources: {source_count}
- Content types: {content_type_summary}

Return the improved paragraph that maintains these facts while improving clarity and professionalism.
LLMB_METADATA_PARAGRAPH_REWRITING_SYSTEM + LLMB_METADATA_PARAGRAPH_REWRITING_USER (31 calls in window)

System prompt

Improve wording and sentence flow only. Preserve every factual value and level of precision; do not convert approximate values into exact ones, infer missing methodology, or add claims about comprehensiveness or quality. Keep the paragraph’s function and roughly its current length. 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 — draft_paragraph: {draft_paragraph}; date_from: {date_from}; date_to: {date_to}; source_count: {source_count}; content_type_summary: {content_type_summary}. Use only these inputs to complete the task defined by the system prompt.