Best LLMs for Short Promotional Teaser Generation
Creates a grounded, editorial-style teaser for a content item from its name, title, source material, and optional source URL. Intended for compact discovery surfaces; returns 1–2 sentences in promotional_text, optionally including the supplied URL and capped at 200 characters in
Models
Frontier on this task: Claude Sonnet 5 at 8.78 / 10. Quality bar at 90%: 7.90.
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 |
|---|---|---|---|---|
| GPT-5.6 Luna | 8.35 / 10 | 8.22 | $0.26 | best value |
| Tencent Hy3 | 8.66 / 10 | 8.55 | $0.82 | 3.2x more expensive |
| NVIDIA Nemotron 3.5 Lightning | 8.05 / 10 | 7.77 | $1.04 | 4.1x more expensive |
| GLM-5.3 Flash | 8.68 / 10 | 8.51 | $1.72 | 6.7x more expensive |
| GPT-5.6 Terra | 8.34 / 10 | 8.22 | $1.80 | 7x more expensive |
| DeepSeek V4 Flash | 8.37 / 10 | 8.14 | $1.81 | 7.1x more expensive |
| NVIDIA Nemotron-3 Super 120B | 8.15 / 10 | 7.99 | $2.50 | 9.7x more expensive |
| GPT-5.6 Sol | 8.22 / 10 | 8.05 | $3.08 | 12x more expensive |
| NVIDIA Nemotron-3 Ultra 550B | 8.64 / 10 | 8.47 | $4.46 | 17x more expensive |
| DeepSeek V4 Pro | 8.59 / 10 | 8.37 | $5.07 | 20x more expensive |
| Gemini 3.5 Flash | 8.02 / 10 | 7.86 | $6.16 | 24x more expensive |
| Qwen 3.7 Plus | 8.37 / 10 | 8.24 | $6.37 | 25x more expensive |
| Claude Sonnet 5 | 8.78 / 10 | 8.67 | $7.05 | 28x more expensive |
| Meta Muse Spark 1.3 | 8.08 / 10 | 7.74 | $8.17 | 32x more expensive |
| Thinking Machines Inkling Small | 8.51 / 10 | 8.37 | $9.16 | 36x more expensive |
| Claude Opus 5 | 8.61 / 10 | 8.41 | $9.61 | 38x more expensive |
| GLM-5.3 | 8.71 / 10 | 8.52 | $16.75 | 65x more expensive |
| Moonshot Kimi K3 | 8.77 / 10 | 8.62 | $21.66 | 85x more expensive |
| Tencent Hy4 Preview | 7.92 / 10 | 7.49 | $23.63 | 92x more expensive |
| Grok 4.6 | 8.24 / 10 | 7.99 | $27.13 | 106x more expensive |
| Thinking Machines Inkling | 8.47 / 10 | 8.27 | $28.70 | 112x more expensive |
| Gemini 3.1 Flash Lite | 7.50 / 10 | 7.28 | $0.35 | 1.4x more expensive |
| Gemini 3.5 Flash Lite | 7.37 / 10 | 7.15 | $0.28 | 1.1x more expensive |
| Gemini 3.8 Flash | 7.81 / 10 | 7.46 | $2.69 | 11x more expensive |
| Qwen 3.8 Max | 7.71 / 10 | 7.35 | $25.53 | 100x more expensive |
| Qwen 3.8 Flash | 7.80 / 10 | 7.52 | $1.53 | 6x more expensive |
| NVIDIA Nemotron-3 Nano 30B-A3B | 7.78 / 10 | 7.46 | $0.47 | 1.8x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| GPT-5.6 Luna ★ OpenAI | 8.35 / 10 CI [8.22, 8.48] | RANKED | $0.26 | best value | batch |
| Tencent Hy3 OpenRouter | 8.66 / 10 CI [8.55, 8.77] | RANKED | $0.82 | 3.2x | batch |
| NVIDIA Nemotron 3.5 Lightning OpenRouter | 8.05 / 10 CI [7.77, 8.34] | HIGH | $1.04 | 4.1x | batch |
| GLM-5.3 Flash Z.AI | 8.68 / 10 CI [8.51, 8.84] | RANKED | $1.72 | 6.7x | batch |
| GPT-5.6 Terra OpenAI | 8.34 / 10 CI [8.22, 8.46] | RANKED | $1.80 | 7x | batch |
| DeepSeek V4 Flash DeepSeek | 8.37 / 10 CI [8.14, 8.61] | HIGH | $1.81 | 7.1x | batch |
| NVIDIA Nemotron-3 Super 120B OpenRouter | 8.15 / 10 CI [7.99, 8.31] | RANKED | $2.50 | 9.7x | batch |
| GPT-5.6 Sol OpenAI | 8.22 / 10 CI [8.05, 8.38] | RANKED | $3.08 | 12x | batch |
| NVIDIA Nemotron-3 Ultra 550B OpenRouter | 8.64 / 10 CI [8.47, 8.81] | RANKED | $4.46 | 17x | batch |
| DeepSeek V4 Pro DeepSeek | 8.59 / 10 CI [8.37, 8.82] | HIGH | $5.07 | 20x | batch |
| Gemini 3.5 Flash Gemini | 8.02 / 10 CI [7.86, 8.19] | RANKED | $6.16 | 24x | batch |
| Qwen 3.7 Plus Alibaba Cloud (DashScope) | 8.37 / 10 CI [8.24, 8.50] | RANKED | $6.37 | 25x | batch |
| Claude Sonnet 5 best Anthropic | 8.78 / 10 CI [8.67, 8.89] | RANKED | $7.05 | 28x | batch |
| Meta Muse Spark 1.3 OpenRouter | 8.08 / 10 CI [7.74, 8.42] | MEDIUM | $8.17 | 32x | batch |
| Thinking Machines Inkling Small OpenRouter | 8.51 / 10 CI [8.37, 8.64] | RANKED | $9.16 | 36x | batch |
| Claude Opus 5 Anthropic | 8.61 / 10 CI [8.41, 8.82] | HIGH | $9.61 | 38x | batch |
| GLM-5.3 Z.AI | 8.71 / 10 CI [8.52, 8.90] | RANKED | $16.75 | 65x | batch |
| Moonshot Kimi K3 Moonshot AI | 8.77 / 10 CI [8.62, 8.93] | RANKED | $21.66 | 85x | batch |
| Tencent Hy4 Preview OpenRouter | 7.92 / 10 CI [7.49, 8.35] | MEDIUM | $23.63 | 92x | batch |
| Grok 4.6 xAI | 8.24 / 10 CI [7.99, 8.50] | HIGH | $27.13 | 106x | batch |
| Thinking Machines Inkling OpenRouter | 8.47 / 10 CI [8.27, 8.67] | RANKED | $28.70 | 112x | 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: 1707 input tokens → 2070 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 grounding and factual accuracy, specificity of the value signal, ability to create legitimate curiosity, natural attribution, concise editorial style, and sensible use or omission of the supplied source URL. If included, the URL must be reproduced exactly and must not displace the teaser’s useful content. Do not assume that the item is a financial report, client publication, or news article unless the input says so. Character count and schema validity are deterministic and out of scope for the semantic judge.
Prompt templates
This is a pooled capability — 2 prompt families share it. The pair shown first is the most frequently used in production.
ACTIVITY_PROMO_SYSTEM_PROMPT +
ACTIVITY_PROMO_USER_PROMPT
(449 calls in window)
System prompt
You are an expert copywriter for a financial research platform. Your task is to write a short, engaging promotional blurb for an activity feed.
The blurb should:
- Be 1-2 sentences, under 200 characters total
- Spark curiosity and encourage clicking through
- Reference the client/brand name naturally
- Feel like an editorial teaser, not an advertisement
- Avoid clickbait, hyperbole, or exclamation marks
- Use present tense
Output your result 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
Client: {client_name}
Post title: {post_title}
Full report text:
{report_text}
Write a short promotional blurb for this post to appear in the activity feed of other clients' pages.
The required JSON output schema is provided in the system prompt.LLMB_SHORT_PROMOTIONAL_TEASER_GENERATION_SYSTEM +
LLMB_SHORT_PROMOTIONAL_TEASER_GENERATION_USER
(201 calls in window)
System prompt
You are an expert editorial copywriter. Write a short, engaging promotional teaser for a content item. The teaser is intended for a compact discovery surface such as an activity feed, content card, digest, or link preview. Use only the supplied title and source material. Surface one specific idea, finding, benefit, tension, or question that gives a reader a reason to open the item. Mention the source, publisher, creator, organization, or brand naturally when useful. Use present tense and an editorial rather than advertising tone. Avoid clickbait, hyperbole, unsupported claims, filler, hashtags, exclamation marks, and generic praise. The user may supply a source URL. Include that URL exactly as supplied only when a direct link is sensible for the teaser and still leaves room for useful copy; otherwise omit it. Never invent, shorten, rewrite, decorate, or add tracking parameters to a URL. Any included URL counts toward the 200-character total. Return 1–2 sentences, no more than 200 characters. 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
Source, publisher, creator, or brand: {client_name}. Item title: {post_title}. Source URL (optional): {source_url}. Source material: {report_text}. Write one short promotional teaser suitable for a compact discovery surface and ground it only in the supplied material.