Best LLMs for Public Engagement Reply Review
Reviews a caller-supplied public response against its source message, allowed evidence, forbidden claims, operator instructions, channel rules, and review policy, returning an advisory readiness verdict and traceable defects without publishing or mutating approval state. Illustra
Models
Frontier on this task: GLM-5.3 Flash at 9.48 / 10. Quality bar at 90%: 8.53.
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 |
|---|---|---|---|---|
| GLM-5.3 Flash | 9.48 / 10 | 9.00 | $2.37 | best value |
| Claude Sonnet 5 | 8.61 / 10 | 8.29 | $6.44 | 2.7x more expensive |
| Tencent Hy4 Preview | 9.11 / 10 | 8.61 | $25.49 | 11x more expensive |
| Claude Haiku 4.5 | 7.14 / 10 | 6.69 | $1.79 | 25% cheaper |
| Tencent Hy3 | 7.95 / 10 | 7.52 | $0.33 | 86% cheaper |
| GPT-5.6 Terra | 8.31 / 10 | 7.97 | $3.17 | 1.3x more expensive |
| GPT-5.6 Luna | 7.97 / 10 | 7.56 | $0.31 | 87% cheaper |
| MiniMax M3 | 7.34 / 10 | 6.90 | $0.97 | 59% cheaper |
| GPT-5.6 Sol | 8.46 / 10 | 8.07 | $5.43 | 2.3x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| GLM-5.3 Flash ★ best Z.AI | 9.48 / 10 CI [9.00, 9.96] | MEDIUM | $2.37 | best value | batch |
| Claude Sonnet 5 Anthropic | 8.61 / 10 CI [8.29, 8.93] | MEDIUM | $6.44 | 2.7x | batch |
| Tencent Hy4 Preview OpenRouter | 9.11 / 10 CI [8.61, 9.61] | MEDIUM | $25.49 | 11x | 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: 1177 input tokens → 1914 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.