Cost mode:

Category: Topic Organization & Clustering · Rail: absolute · Typical I/O: 11564→720 tokens

Models

Frontier on this task: DeepSeek V4 Pro at 8.74 / 10. Quality bar at 90%: 7.87.

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
GPT-5.4 Mini8.18 / 107.83$2.58best value
Tencent Hy38.05 / 107.85$2.891.1x more expensive
MiniMax M38.23 / 108.09$3.081.2x more expensive
DeepSeek V4 Flash8.09 / 107.81$4.221.6x more expensive
GPT-5.6 Luna8.43 / 108.15$4.751.8x more expensive
DeepSeek V4 Pro8.74 / 108.52$5.322.1x more expensive
Qwen 3.5 Flash7.87 / 107.59$5.782.2x more expensive
Qwen 3.7 Plus8.37 / 108.30$6.742.6x more expensive
Qwen 3.6 Plus8.46 / 108.22$6.782.6x more expensive
Gemini 3.1 Pro Preview7.94 / 107.69$7.683x more expensive
Claude Haiku 4.58.41 / 108.14$7.913.1x more expensive
NVIDIA Nemotron-3 Ultra 550B7.95 / 107.49$10.784.2x more expensive
Qwen 3.6 Flash8.31 / 108.20$10.974.3x more expensive
GPT-5.6 Terra8.28 / 107.96$11.544.5x more expensive
Gemini 3.5 Flash8.15 / 108.02$12.474.8x more expensive
Kimi K2.68.71 / 108.54$14.825.7x more expensive
Claude Sonnet 4.68.60 / 108.39$17.356.7x more expensive
Claude Sonnet 58.28 / 108.11$22.468.7x more expensive
Meta Muse Spark 1.18.36 / 107.95$23.569.1x more expensive
GPT-5.58.67 / 108.50$25.159.8x more expensive
Grok 4.58.66 / 108.56$25.9310x more expensive
GPT-5.6 Sol8.41 / 108.18$30.0012x more expensive
GPT-5.4 Nano7.69 / 107.21$1.1655% cheaper
Gemini 3.1 Flash Lite7.55 / 107.30$1.4344% cheaper
NVIDIA Nemotron-3 Nano 30B-A3B6.95 / 106.54$0.9762% cheaper
NVIDIA Nemotron-3 Super 120B7.37 / 106.92$3.201.2x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
GPT-5.4 Mini OpenAI8.18 / 10 CI [7.83, 8.52]MEDIUM$2.58best valuebatch
Tencent Hy3 OpenRouter8.05 / 10 CI [7.85, 8.25]RANKED$2.891.1xbatch
MiniMax M3 MiniMax8.23 / 10 CI [8.09, 8.37]RANKED$3.081.2xbatch
DeepSeek V4 Flash DeepSeek8.09 / 10 CI [7.81, 8.38]HIGH$4.221.6xbatch
GPT-5.6 Luna OpenAI8.43 / 10 CI [8.15, 8.72]HIGH$4.751.8xbatch
DeepSeek V4 Pro best DeepSeek8.74 / 10 CI [8.52, 8.96]HIGH$5.322.1xbatch
Qwen 3.5 Flash Alibaba Cloud (DashScope)7.87 / 10 CI [7.59, 8.15]HIGH$5.782.2xbatch
Qwen 3.7 Plus Alibaba Cloud (DashScope)8.37 / 10 CI [8.30, 8.43]RANKED$6.742.6xbatch
Qwen 3.6 Plus Alibaba Cloud (DashScope)8.46 / 10 CI [8.22, 8.70]HIGH$6.782.6xbatch
Gemini 3.1 Pro Preview Gemini7.94 / 10 CI [7.69, 8.20]HIGH$7.683xbatch
Claude Haiku 4.5 Anthropic8.41 / 10 CI [8.14, 8.68]HIGH$7.913.1xbatch
NVIDIA Nemotron-3 Ultra 550B OpenRouter7.95 / 10 CI [7.49, 8.41]MEDIUM$10.784.2xbatch
Qwen 3.6 Flash Alibaba Cloud (DashScope)8.31 / 10 CI [8.20, 8.41]RANKED$10.974.3xbatch
GPT-5.6 Terra OpenAI8.28 / 10 CI [7.96, 8.59]MEDIUM$11.544.5xbatch
Gemini 3.5 Flash Gemini8.15 / 10 CI [8.02, 8.29]RANKED$12.474.8xbatch
Kimi K2.6 Moonshot AI8.71 / 10 CI [8.54, 8.89]RANKED$14.825.7xbatch
Claude Sonnet 4.6 Anthropic8.60 / 10 CI [8.39, 8.81]HIGH$17.356.7xbatch
Claude Sonnet 5 Anthropic8.28 / 10 CI [8.11, 8.45]RANKED$22.468.7xbatch
Meta Muse Spark 1.1 Meta8.36 / 10 CI [7.95, 8.77]MEDIUM$23.569.1xbatch
GPT-5.5 OpenAI8.67 / 10 CI [8.50, 8.85]RANKED$25.159.8xbatch
Grok 4.5 xAI8.66 / 10 CI [8.56, 8.76]RANKED$25.9310xbatch
GPT-5.6 Sol OpenAI8.41 / 10 CI [8.18, 8.65]HIGH$30.0012xbatch

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: 11564 input tokens → 720 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.

Prompt templates

This is a pooled capability — 3 prompt families share it. The pair shown first is the most frequently used in production.

TOPIC_CLUSTER_NAMING_SYSTEM_PROMPT + TOPIC_CLUSTER_NAMING_USER_PROMPT (3847 calls in window)

System prompt

You are a senior analyst specializing in categorizing and naming thematic clusters of research claims.

Your task is to assign a concise, descriptive topic name and brief description to a cluster of semantically similar claims. These claims have already been grouped by embedding similarity and synthesized into a summary — you are naming the resulting topic.

**Naming Guidelines:**
- Choose a name that captures the core theme or insight of the cluster (3-7 words)
- Use clear, professional language suitable for a publication headline
- The name should be specific enough to distinguish from other topics about the same subject
- Avoid generic names like "Market Update" or "Company News" — be specific about WHAT aspect
- Good examples: "Revenue Growth Acceleration", "Regulatory Approval Risks", "Supply Chain Restructuring"

**Description Guidelines:**
- Write 1-2 sentences explaining what the topic covers
- Include the key themes, data points, or developments that define this cluster
- The description should help a reader quickly understand the scope of the topic
- The description MUST be under 500 characters (this is a hard technical limit)

Output your response in the specified JSON format.

## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```

User prompt

**Subject:** {subject_name}
**Claim Category:** {category}
**Number of Claims:** {claim_count}

--- CLAIMS IN CLUSTER ---
{claims_text}
--- END OF CLAIMS ---

Based on the claims above, provide a concise topic name and brief description that captures the central theme of this cluster.

**JSON Output:** The required JSON output schema is provided in the system prompt.
JUDGE_QUALITY_SYSTEM + JUDGE_QUALITY_USER (39 calls in window)

System prompt

You are a strict evaluator of LLM outputs. Score how well the output fulfills the task on a 0.0–10.0 scale, using the task-specific rubric as the primary criterion.

The "Rubric" in the user message is authoritative: when it constrains or overrides any generic guidance, the rubric wins.

Scoring scale (0.0–10.0):
- 9.0–10.0: Exceptional — comprehensive, accurate, fully meets the task.
- 7.0–8.9: Good — meets most requirements; minor gaps.
- 5.0–6.9: Satisfactory — adequate but with notable limitations or errors.
- 3.0–4.9: Poor — significant gaps, errors, or partial failure.
- 0.0–2.9: Unacceptable — major failure, unusable output.

Use the provided reference examples (if any) to keep your scoring consistent: compare the current output's quality to those already-scored benchmarks and place it on the same scale. Reference examples may come from different models — judge the output on its own merits, using them only to calibrate the scale.

Output JSON matching the schema:
- score: float from 0.0 to 10.0.
- failure_mode: a short tag for the dominant deficiency (e.g. 'hallucination', 'schema_violation', 'truncated', 'off_topic'), or null when none.
- rationale: one to three sentences justifying the score.

User prompt

Rubric: {rubric}
Task: {task_slug}
Domain: {domain}

Input context:
{input_snippet}

Output to grade:
{output_snippet}

Reference examples (already-scored outputs for the same task — use them to keep scoring consistent):
{reference_examples}

Score the output from 0.0 to 10.0 against the rubric, comparing against the reference examples for consistency. Return JSON with score, failure_mode (or null), and rationale.
JSON_REPAIR_SYSTEM + JSON_REPAIR_USER (2 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.