Best LLMs for Engagement Reply Review
Independent quality review of a drafted engagement reply.
Models
Frontier on this task: Qwen 3.6 Plus at 9.88 / 10. Quality bar at 90%: 8.89.
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.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| Gemini 3.1 Flash Lite | 9.80 / 10 | 9.66 | $0.21 | best value |
| Qwen 3.5 Flash | 9.71 / 10 | 9.52 | $1.76 | 8.5x more expensive |
| Claude Sonnet 5 | 8.97 / 10 | 8.62 | $6.29 | 30x more expensive |
| Qwen 3.6 Plus | 9.88 / 10 | 9.79 | $7.94 | 38x more expensive |
| Claude Opus 4.8 | 9.77 / 10 | 9.62 | $12.00 | 58x more expensive |
| Kimi K2.6 | 9.53 / 10 | 9.18 | $17.08 | 83x more expensive |
| GPT-5.4 Nano | 6.98 / 10 | 6.57 | $0.31 | 1.5x more expensive |
| GPT-5.6 Sol | 8.63 / 10 | 8.21 | $7.82 | 38x more expensive |
| GPT-5.6 Luna | 8.09 / 10 | 7.65 | $1.35 | 6.5x more expensive |
| GPT-5.4 Mini | 6.94 / 10 | 6.48 | $0.47 | 2.3x more expensive |
| GPT-5.5 | 7.99 / 10 | 7.68 | $8.71 | 42x more expensive |
| GPT-5.6 Terra | 8.48 / 10 | 8.10 | $3.30 | 16x more expensive |
| Claude Haiku 4.5 | 7.38 / 10 | 6.90 | $1.72 | 8.3x more expensive |
| Claude Sonnet 4.6 | 8.01 / 10 | 7.55 | $6.36 | 31x more expensive |
| Tencent Hy3 | 8.15 / 10 | 7.67 | $0.97 | 4.7x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| Gemini 3.1 Flash Lite ★ Gemini | 9.80 / 10 CI [9.66, 9.94] | RANKED | $0.21 | best value | batch |
| Qwen 3.5 Flash Alibaba Cloud (DashScope) | 9.71 / 10 CI [9.52, 9.91] | RANKED | $1.76 | 8.5x | batch |
| Claude Sonnet 5 Anthropic | 8.97 / 10 CI [8.62, 9.32] | MEDIUM | $6.29 | 30x | batch |
| Qwen 3.6 Plus best Alibaba Cloud (DashScope) | 9.88 / 10 CI [9.79, 9.97] | RANKED | $7.94 | 38x | batch |
| Claude Opus 4.8 Anthropic | 9.77 / 10 CI [9.62, 9.91] | RANKED | $12.00 | 58x | batch |
| Kimi K2.6 Moonshot AI | 9.53 / 10 CI [9.18, 9.89] | MEDIUM | $17.08 | 83x | 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: 1029 input tokens → 961 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
This capability is itself a REVIEWER: given a social post and a drafted public reply, it outputs a verdict — a 0..1 score, a pass/fail decision, and reasoning — on whether the reply is good enough to post for llm-bench's community engagement. You are judging the QUALITY OF THAT REVIEW VERDICT — not the reply itself, and not whether the JSON is well-formed. A syntactically perfect, confident-sounding verdict that reached the wrong decision or mischaracterised the reply is a POOR output, not a good one. First, independently assess the drafted reply yourself against the reviewer's own criteria: - Genuinely helpful and on-topic for THIS post — would it stand as a useful answer with the brand name and tracked link removed? If it is off-topic, boilerplate, or brand-recitation, the correct verdict is score <= 0.3 and passed=false. - Natural voice, NOT salesy or spammy. "Salesy" means describing what llm-bench is or does, repeating the brand, or a generic pitch detached from the post. A single earned mention carrying a first-party kapualabs.com / post.kapualabs.com tracked link is acceptable — do not treat that domain as third-party spam. - Any llm-bench mention is earned, not forced. - Claim-safe: no forbidden claims, nothing unverifiable. - Appropriate for the thread. Then judge whether the review's decision AND its reasoning were correct: - Did its score / passed decision match your own assessment? Heavily penalise a verdict that APPROVED an off-topic, spammy, promotional, boilerplate, or forbidden-claim reply, or that REJECTED a genuinely good one. - Did its reasoning accurately describe the reply? Heavily penalise a verdict that MISCHARACTERISED the draft — e.g. called jargon-heavy text "conversational", claimed the reply is "lightly funny" when it contains no humour, or asserted a relevance that is not there. A hallucinated or self-contradicting characterisation is a serious error even when the final number happens to be defensible. Reward verdicts that reach the correct decision for the right reasons and describe the reply accurately. Do NOT award a high score merely because the output follows the JSON schema, is well-formatted, or sounds confident — grade the substance of the judgement, not its surface form. Use the FULL 0-10 range for YOUR score (this is your grade of the review, distinct from the review's own 0..1 score described above): - 9-10: the verdict reached the correct decision AND its reasoning accurately describes the reply. Reserve this for genuinely flawless reviews. - 6-8: broadly correct decision but with a real weakness — a shallow or partly inaccurate rationale, or a score that is defensible but not well-justified. - 3-5: the reasoning materially mischaracterises the reply (e.g. hallucinated tone/humour, asserted relevance that is not there), even if the final number is arguably defensible. - 0-2: the verdict is simply wrong — it approved an off-topic, spammy, promotional, boilerplate, or forbidden-claim reply, or rejected a genuinely good one, however well-formatted it is. If you find yourself placing most reviews between 9 and 10, you are grading surface form rather than substance — go back and look harder for the errors above. Do not inflate a review that got the number right by luck while reasoning poorly.
Prompt templates
The system + user template pair used for this task.
JSON_REPAIR_SYSTEM +
JSON_REPAIR_USER
(39 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.