Fronset Benchmark

Best LLMs for Research Query Generation

Generates diverse, high-signal research queries for a subject, purpose, region, and target search platform while respecting platform syntax and caller requirements. Illustrative uses include generating searches for vendor diligence, market entry, software vulnerabilities, litigat

Cost mode:

Category: Infrastructure & Utility · Typical I/O: 3424→735 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: GLM-5.3 Flash at 7.87 / 10. Quality bar at 90%: 7.08.

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.

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
GPT-6 Luna ★ OpenAI7.09 / 10 CI [6.83, 7.36]HIGH$0.22best valuebatch
Xiaomi MiMo V2.6 Pro Xiaomi MiMo7.51 / 10 CI [7.26, 7.75]HIGH$0.502.3xbatch
Gemini 3.8 Flash Gemini7.66 / 10 CI [7.41, 7.90]HIGH$3.7917xbatch
Thinking Machines Inkling Small OpenRouter7.36 / 10 CI [7.02, 7.70]MEDIUM$4.0519xbatch
GPT-6 Sol OpenAI7.20 / 10 CI [6.88, 7.53]MEDIUM$4.2119xbatch
GLM-5.3 Flash best Z.AI7.87 / 10 CI [7.67, 8.07]HIGH$5.2124xbatch
Claude Sonnet 5.5 Anthropic7.17 / 10 CI [6.79, 7.54]MEDIUM$6.3629xbatch
Claude Opus 5.5 Anthropic7.24 / 10 CI [6.82, 7.66]MEDIUM$11.8755xbatch
Meta Muse Spark 1.3 OpenRouter7.62 / 10 CI [7.34, 7.91]HIGH$13.6863xbatch
Grok 4.7 xAI7.43 / 10 CI [7.15, 7.71]HIGH$18.1984xbatch
Tencent Hy4 Preview OpenRouter7.34 / 10 CI [7.13, 7.55]HIGH$21.99101xbatch
Moonshot Kimi K3 Moonshot AI7.80 / 10 CI [7.50, 8.10]MEDIUM$29.09134xbatch
Thinking Machines Inkling OpenRouter7.34 / 10 CI [6.94, 7.74]MEDIUM$35.18162xbatch
GLM-5.3 Z.AI7.73 / 10 CI [7.49, 7.96]HIGH$57.02262xbatch

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: 3424 input tokens → 735 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 purpose coverage, expected search utility, diversity, subject and regional relevance, platform appropriateness, and absence of unsupported assumptions. Penalize near-duplicates and over-constrained queries likely to suppress useful results.

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.

{
  "$defs": {
    "ResearchQuery": {
      "description": "Individual research query",
      "properties": {
        "priority": {
          "description": "Query priority (lower = higher priority)",
          "title": "Priority",
          "type": "integer"
        },
        "query_text": {
          "description": "The search query text",
          "title": "Query Text",
          "type": "string"
        },
        "query_type": {
          "description": "Type of query (general, news, etc.)",
          "title": "Query Type",
          "type": "string"
        },
        "reasoning": {
          "description": "Reasoning for this query",
          "title": "Reasoning",
          "type": "string"
        }
      },
      "required": [
        "query_text",
        "query_type",
        "priority",
        "reasoning"
      ],
      "title": "ResearchQuery",
      "type": "object"
    }
  },
  "description": "Output model for query generation",
  "properties": {
    "queries": {
      "description": "List of generated queries",
      "items": {
        "$ref": "#/$defs/ResearchQuery"
      },
      "title": "Queries",
      "type": "array"
    }
  },
  "required": [
    "queries"
  ],
  "title": "QueryGenerationOutput",
  "type": "object"
}

Prompt templates

The published template for this task.

default · LLMB_RESEARCH_QUERY_GENERATION_SYSTEM + LLMB_RESEARCH_QUERY_GENERATION_USER (0 calls in window)

System prompt

Generate queries that collectively cover the supplied research purpose and chapter context.

Target platform rules:
{platform_rules}
 Use the target platform’s supported syntax, terminology real sources are likely to use, and region or language context when relevant. Vary facets rather than paraphrasing one query. Do not add unsupported dates, companies, tickers, or operators; leave final syntax enforcement to validation. 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 — num_queries: {num_queries}; subject_name: {subject_name}; subject_code: {subject_code}; search_platform: {search_platform}; target_region: {target_region}; subject_type: {subject_type}; subject_description: {subject_description}; chapter_context: {chapter_context}; query_purpose: {query_purpose}; additional_requirements: {additional_requirements}. Use only these inputs to complete the task defined by the system prompt.