Best LLMs for Metadata Paragraph Rewriting

Category: Infrastructure & Utility · Typical I/O: 587→1582 tokens

Runs on model group: All models. Inputs may reach any model in the group, not only the ranked models below.

Run this task on Fronset — request an invitation

Models

Frontier on this task: NVIDIA Nemotron-3 Ultra 550B at 8.36 / 10. Quality bar at 90%: 7.52.

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 Flash	7.81 / 10	7.60	$0.06	best value
GPT-6 Luna	7.81 / 10	7.63	$0.08	1.4x more expensive
Claude Haiku 5.5	7.95 / 10	7.72	$0.10	1.6x more expensive
Xiaomi MiMo V2.6 Pro	7.89 / 10	7.71	$0.23	3.9x more expensive
MiniMax M3	7.52 / 10	7.04	$0.35	5.9x more expensive
DeepSeek V4.1 Flash	7.90 / 10	7.73	$0.61	10x more expensive
Claude Sonnet 5.5	7.80 / 10	7.64	$1.30	22x more expensive
GPT-6 Sol	8.00 / 10	7.84	$1.71	29x more expensive
NVIDIA Nemotron-3 Ultra 550B	8.36 / 10	8.10	$1.87	32x more expensive
Gemini 3.8 Flash	7.64 / 10	7.42	$3.06	52x more expensive
Claude Opus 5.5	8.00 / 10	7.75	$3.70	63x more expensive
Meta Muse Spark 1.3	7.54 / 10	7.16	$7.02	119x more expensive
Grok 4.7	7.84 / 10	7.69	$13.66	232x more expensive
Qwen 3.8 Max	7.59 / 10	7.36	$15.28	259x more expensive
NVIDIA Nemotron-3 Super 120B	7.42 / 10	7.09	$0.40	6.7x more expensive
Tencent Hy3	6.93 / 10	6.74	$0.17	2.8x more expensive
Thinking Machines Inkling	6.76 / 10	6.58	$6.16	105x more expensive
Gemini 3.5 Flash Lite	6.87 / 10	6.76	$0.20	3.5x more expensive
NVIDIA Nemotron 3.5 Lightning	7.30 / 10	6.97	$0.82	14x more expensive
Qwen 3.8 Flash	7.46 / 10	7.17	$2.16	37x more expensive
Tencent Hy4 Preview	7.36 / 10	7.04	$26.66	453x more expensive
Thinking Machines Inkling Small	6.91 / 10	6.60	$1.54	26x more expensive
NVIDIA Nemotron-3 Nano 30B-A3B	6.80 / 10	6.51	$0.20	3.4x more expensive
Qwen 3.7 Plus	7.38 / 10	7.16	$2.92	50x more expensive
Moonshot Kimi K3	7.39 / 10	7.14	$19.08	324x more expensive

Cost breakdown

Model	Quality	Confidence	Cost / 1k runs	Overpay	Mode
Xiaomi MiMo V2.6 Flash ★ Xiaomi MiMo	7.81 / 10 CI [7.60, 8.03]	HIGH	$0.06	best value	batch
GPT-6 Luna OpenAI	7.81 / 10 CI [7.63, 7.99]	RANKED	$0.08	1.4x	batch
Claude Haiku 5.5 Anthropic	7.95 / 10 CI [7.72, 8.18]	HIGH	$0.10	1.6x	batch
Xiaomi MiMo V2.6 Pro Xiaomi MiMo	7.89 / 10 CI [7.71, 8.07]	RANKED	$0.23	3.9x	batch
MiniMax M3 OpenRouter	7.52 / 10 CI [7.04, 8.01]	MEDIUM	$0.35	5.9x	batch
DeepSeek V4.1 Flash DeepSeek	7.90 / 10 CI [7.73, 8.07]	RANKED	$0.61	10x	batch
Claude Sonnet 5.5 Anthropic	7.80 / 10 CI [7.64, 7.96]	RANKED	$1.30	22x	batch
GPT-6 Sol OpenAI	8.00 / 10 CI [7.84, 8.17]	RANKED	$1.71	29x	batch
NVIDIA Nemotron-3 Ultra 550B best OpenRouter	8.36 / 10 CI [8.10, 8.62]	HIGH	$1.87	32x	batch
Gemini 3.8 Flash Gemini	7.64 / 10 CI [7.42, 7.86]	HIGH	$3.06	52x	batch
Claude Opus 5.5 Anthropic	8.00 / 10 CI [7.75, 8.25]	HIGH	$3.70	63x	batch
Meta Muse Spark 1.3 OpenRouter	7.54 / 10 CI [7.16, 7.92]	MEDIUM	$7.02	119x	batch
Grok 4.7 xAI	7.84 / 10 CI [7.69, 7.99]	RANKED	$13.66	232x	batch
Qwen 3.8 Max Alibaba Cloud (DashScope)	7.59 / 10 CI [7.36, 7.83]	HIGH	$15.28	259x	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: 587 input tokens → 1582 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.

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.

{
  "description": "Defines the structure for LLM-improved metadata paragraph about research period and sources.",
  "properties": {
    "improved_paragraph": {
      "description": "The improved, professionally worded metadata paragraph describing the research period, source count, and content types. Should be 2-3 sentences, clear, concise, and suitable for inclusion at the beginning of a research report.",
      "title": "Improved Paragraph",
      "type": "string"
    }
  },
  "required": [
    "improved_paragraph"
  ],
  "title": "ImprovedMetadataParagraph",
  "type": "object"
}

Prompt templates

The published template for this task.

default · LLMB_METADATA_PARAGRAPH_REWRITING_SYSTEM + LLMB_METADATA_PARAGRAPH_REWRITING_USER (96 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.