Cost mode:

Category: Topic Organization & Clustering · Rail: absolute · Typical I/O: 22325→1424 tokens

Models

Frontier on this task: Claude Opus 5 at 8.91 / 10. Quality bar at 90%: 8.02.

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
MiniMax M38.21 / 108.06$3.19best value
Gemini 3.8 Flash8.24 / 107.97$5.771.8x more expensive
Tencent Hy38.14 / 107.96$6.942.2x more expensive
Thinking Machines Inkling Small8.34 / 107.96$9.182.9x more expensive
Claude Haiku 4.58.03 / 107.64$9.212.9x more expensive
Qwen 3.7 Plus8.09 / 107.67$9.553x more expensive
Gemini 3.5 Flash8.22 / 108.08$13.914.4x more expensive
DeepSeek V4 Flash8.12 / 107.90$14.064.4x more expensive
DeepSeek V4 Pro8.45 / 108.17$20.486.4x more expensive
Thinking Machines Inkling8.39 / 107.97$22.867.2x more expensive
GPT-5.6 Sol8.34 / 108.03$23.117.2x more expensive
Claude Sonnet 58.37 / 108.19$26.548.3x more expensive
Tencent Hy4 Preview8.40 / 107.91$36.5711x more expensive
Meta Muse Spark 1.38.11 / 107.76$36.5811x more expensive
GLM-5.38.72 / 108.48$37.3312x more expensive
Claude Opus 58.91 / 108.58$42.1513x more expensive
Grok 4.68.57 / 108.35$48.0315x more expensive
Moonshot Kimi K38.91 / 108.74$64.2320x more expensive
Qwen 3.8 Flash7.57 / 107.10$3.661.1x more expensive
NVIDIA Nemotron 3.5 Lightning7.28 / 106.78$2.5221% cheaper
Qwen 3.8 Max7.96 / 107.55$59.0318x more expensive
NVIDIA Nemotron-3 Nano 30B-A3B6.93 / 106.49$1.6349% cheaper
GPT-5.6 Luna8.00 / 107.68$3.251x more expensive
Gemini 3.5 Flash Lite7.35 / 107.02$2.4324% cheaper
Gemini 3.1 Flash Lite7.65 / 107.43$2.0436% cheaper
GPT-5.4 Nano7.31 / 106.87$1.8343% cheaper
GPT-5.6 Terra7.87 / 107.43$11.103.5x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
MiniMax M3 OpenRouter8.21 / 10 CI [8.06, 8.37]RANKED$3.19best valuebatch
Gemini 3.8 Flash Gemini8.24 / 10 CI [7.97, 8.51]HIGH$5.771.8xbatch
Tencent Hy3 OpenRouter8.14 / 10 CI [7.96, 8.33]RANKED$6.942.2xbatch
Thinking Machines Inkling Small OpenRouter8.34 / 10 CI [7.96, 8.71]MEDIUM$9.182.9xbatch
Claude Haiku 4.5 Anthropic8.03 / 10 CI [7.64, 8.43]MEDIUM$9.212.9xbatch
Qwen 3.7 Plus Alibaba Cloud (DashScope)8.09 / 10 CI [7.67, 8.50]MEDIUM$9.553xbatch
Gemini 3.5 Flash Gemini8.22 / 10 CI [8.08, 8.35]RANKED$13.914.4xbatch
DeepSeek V4 Flash DeepSeek8.12 / 10 CI [7.90, 8.35]HIGH$14.064.4xbatch
DeepSeek V4 Pro DeepSeek8.45 / 10 CI [8.17, 8.74]HIGH$20.486.4xbatch
Thinking Machines Inkling OpenRouter8.39 / 10 CI [7.97, 8.81]MEDIUM$22.867.2xbatch
GPT-5.6 Sol OpenAI8.34 / 10 CI [8.03, 8.64]MEDIUM$23.117.2xbatch
Claude Sonnet 5 Anthropic8.37 / 10 CI [8.19, 8.56]RANKED$26.548.3xbatch
Tencent Hy4 Preview OpenRouter8.40 / 10 CI [7.91, 8.89]MEDIUM$36.5711xbatch
Meta Muse Spark 1.3 OpenRouter8.11 / 10 CI [7.76, 8.47]MEDIUM$36.5811xbatch
GLM-5.3 Z.AI8.72 / 10 CI [8.48, 8.95]HIGH$37.3312xbatch
Claude Opus 5 best Anthropic8.91 / 10 CI [8.58, 9.24]MEDIUM$42.1513xbatch
Grok 4.6 xAI8.57 / 10 CI [8.35, 8.79]HIGH$48.0315xbatch
Moonshot Kimi K3 Moonshot AI8.91 / 10 CI [8.74, 9.08]RANKED$64.2320xbatch

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: 22325 input tokens → 1424 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 representativeness, specificity, distinction from neighboring concepts, support across cluster members, label clarity, concise boundary description, and absence of unsupported interpretation.

Prompt templates

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

LLMB_TOPIC_CLUSTER_LABELING_SYSTEM + LLMB_TOPIC_CLUSTER_LABELING_USER (1671 calls in window)

System prompt

Name the shared substantive idea that distinguishes this cluster from adjacent clusters. The name must be reusable across subjects: state the theme in wording that a cluster about a different company, market or programme could match into, and keep company, organisation, product, ticker and person names out of it — subject_name is context for reading the claims, never a component of the name. The specifics belong in the description: name there the entities, mechanisms, outcomes or tensions supported by multiple members, and state the common scope and the boundary that separates this cluster from its neighbours. Avoid generic labels such as “Overview,” unsupported conclusions, category repetition, and wording derived from a single outlier. 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 — subject_name: {subject_name}; category: {category}; claim_count: {claim_count}; claims_text: {claims_text}. Use only these inputs to complete the task defined by the system prompt.
TOPIC_CLUSTER_NAMING_SYSTEM_PROMPT + TOPIC_CLUSTER_NAMING_USER_PROMPT (438 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.
JSON_REPAIR_SYSTEM + JSON_REPAIR_USER (6 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.