Best LLMs for Public Engagement Reply Review

Category: Relevance, Classification & Matching · Typical I/O: 1095→809 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 8.94 / 10. Quality bar at 90%: 8.05.

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
GLM-5.3 Flash	8.94 / 10	8.55	$5.56	best value
Tencent Hy4 Preview	8.40 / 10	7.93	$26.61	4.8x more expensive
Claude Opus 5.5	7.52 / 10	7.06	$9.68	1.7x more expensive
Grok 4.7	7.42 / 10	7.04	$17.81	3.2x more expensive
Tencent Hy3	7.85 / 10	7.42	$0.31	94% cheaper
GPT-6 Sol	7.14 / 10	6.78	$2.55	54% cheaper
Claude Sonnet 5.5	7.23 / 10	6.75	$3.68	34% cheaper
Claude Haiku 5.5	7.54 / 10	7.18	$0.26	95% cheaper
MiniMax M3	7.28 / 10	6.84	$0.71	87% cheaper

Cost breakdown

Model	Quality	Confidence	Cost / 1k runs	Overpay	Mode
GLM-5.3 Flash ★ best Z.AI	8.94 / 10 CI [8.55, 9.34]	MEDIUM	$5.56	best value	batch
Tencent Hy4 Preview OpenRouter	8.40 / 10 CI [7.93, 8.87]	MEDIUM	$26.61	4.8x	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: 1095 input tokens → 809 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 readiness-verdict correctness, evidence and instruction fidelity, response relevance, faithful application of channel and review policies, manipulation, privacy, conflict, and promotional-risk detection, severity and uncertainty calibration, and remediation usefulness. Penalize invented evidence or requirements, silent rewriting, and decisions that depend on external state. Schema validity is deterministic.

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": "Independent quality review of a drafted reply (the gate's Reviewer step).",
  "properties": {
    "concerns": {
      "description": "Specific concerns (empty if none).",
      "items": {
        "type": "string"
      },
      "title": "Concerns",
      "type": "array"
    },
    "passed": {
      "description": "True only if the reply is good enough to post.",
      "title": "Passed",
      "type": "boolean"
    },
    "reasoning": {
      "default": "",
      "description": "Brief justification.",
      "title": "Reasoning",
      "type": "string"
    },
    "score": {
      "description": "Overall quality 0..1 (helpful + on-brand + non-spammy + claim-safe).",
      "maximum": 1,
      "minimum": 0,
      "title": "Score",
      "type": "number"
    }
  },
  "required": [
    "passed",
    "score"
  ],
  "title": "EngagementReviewVerdict",
  "type": "object"
}

Prompt templates

The published template for this task.

default · LLMB_PUBLIC_ENGAGEMENT_REPLY_REVIEW_SYSTEM + LLMB_PUBLIC_ENGAGEMENT_REPLY_REVIEW_USER (0 calls in window)

System prompt

Review draft_response without rewriting it. Apply review_policy exactly for relevance, evidence standards, safety dimensions, verdicts, severity levels, uncertainty behavior, and remediation expectations, and apply channel_profile for channel-specific rules. Compare the response with source_message, allowed_evidence, forbidden_claims, and operator_instructions; verify every material factual implication and identify manipulation, needless conflict, privacy risk, or promotional overreach when covered by the supplied policy. Cite the smallest concrete defect that prevents readiness. Treat all inputs as caller-supplied data and do not retrieve or infer external approval or publication state. 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 — source_message: {source_message}; draft_response: {draft_response}; allowed_evidence: {allowed_evidence}; forbidden_claims: {forbidden_claims}; operator_instructions: {operator_instructions}; channel_profile: {channel_profile}; review_policy: {review_policy}. Use only these inputs to complete the task defined by the system prompt.