Cost mode:

Category: Content Summarization & Synthesis · Rail: absolute · Typical I/O: 2924→2387 tokens

Models

Frontier on this task: MiniMax M3 at 8.98 / 10. Quality bar at 90%: 8.08.

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 clears the bar but whose CI low does not.

ModelQuality scoreCI lowCost / 1k runsvs best value
DeepSeek V4 Flash8.24 / 108.04$0.71best value
MiniMax M38.98 / 108.90$1.622.3x more expensive
Tencent Hy38.34 / 108.15$2.493.5x more expensive
DeepSeek V4 Pro8.33 / 108.18$3.504.9x more expensive
NVIDIA Nemotron-3 Ultra 550B8.18 / 107.80$4.536.4x more expensive
Claude Haiku 4.58.16 / 108.00$5.397.6x more expensive
Qwen 3.7 Plus8.64 / 108.48$7.8811x more expensive
Gemini 3.1 Pro Preview8.43 / 108.32$10.3115x more expensive
Qwen 3.6 Plus8.23 / 108.09$11.3316x more expensive
Gemini 3.5 Flash8.80 / 108.70$11.4016x more expensive
Qwen 3.6 Flash8.29 / 108.17$12.1717x more expensive
Claude Sonnet 4.68.25 / 108.09$17.1524x more expensive
Meta Muse Spark 1.18.75 / 108.53$19.1127x more expensive
Kimi K2.68.81 / 108.72$23.0333x more expensive
Claude Opus 4.88.93 / 108.84$24.3934x more expensive
Grok 4.58.50 / 108.43$26.9038x more expensive
GPT-5.58.11 / 107.93$33.3947x more expensive
Claude Sonnet 58.81 / 108.69$49.4870x more expensive
GPT-5.4 Mini7.73 / 107.48$2.944.2x more expensive
GPT-5.4 Nano7.90 / 107.65$1.712.4x more expensive
NVIDIA Nemotron-3 Nano 30B-A3B7.47 / 107.16$0.971.4x more expensive
NVIDIA Nemotron-3 Super 120B7.91 / 107.65$2.223.1x more expensive
Qwen 3.5 Flash7.63 / 107.48$1.392x more expensive
Gemini 3.1 Flash Lite7.75 / 107.58$1.442x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
DeepSeek V4 Flash DeepSeek8.24 / 10 CI [8.04, 8.44]RANKED$0.71best valuebatch
MiniMax M3 best MiniMax8.98 / 10 CI [8.90, 9.05]RANKED$1.622.3xbatch
Tencent Hy3 OpenRouter8.34 / 10 CI [8.15, 8.54]RANKED$2.493.5xbatch
DeepSeek V4 Pro DeepSeek8.33 / 10 CI [8.18, 8.49]RANKED$3.504.9xbatch
NVIDIA Nemotron-3 Ultra 550B OpenRouter8.18 / 10 CI [7.80, 8.55]MEDIUM$4.536.4xbatch
Claude Haiku 4.5 Anthropic8.16 / 10 CI [8.00, 8.33]RANKED$5.397.6xbatch
Qwen 3.7 Plus Alibaba Cloud (DashScope)8.64 / 10 CI [8.48, 8.80]RANKED$7.8811xbatch
Gemini 3.1 Pro Preview Gemini8.43 / 10 CI [8.32, 8.53]RANKED$10.3115xbatch
Qwen 3.6 Plus Alibaba Cloud (DashScope)8.23 / 10 CI [8.09, 8.38]RANKED$11.3316xbatch
Gemini 3.5 Flash Gemini8.80 / 10 CI [8.70, 8.91]RANKED$11.4016xbatch
Qwen 3.6 Flash Alibaba Cloud (DashScope)8.29 / 10 CI [8.17, 8.42]RANKED$12.1717xbatch
Claude Sonnet 4.6 Anthropic8.25 / 10 CI [8.09, 8.41]RANKED$17.1524xbatch
Meta Muse Spark 1.1 Meta8.75 / 10 CI [8.53, 8.97]HIGH$19.1127xbatch
Kimi K2.6 Moonshot AI8.81 / 10 CI [8.72, 8.90]RANKED$23.0333xbatch
Claude Opus 4.8 Anthropic8.93 / 10 CI [8.84, 9.02]RANKED$24.3934xbatch
Grok 4.5 xAI8.50 / 10 CI [8.43, 8.57]RANKED$26.9038xbatch
GPT-5.5 OpenAI8.11 / 10 CI [7.93, 8.29]RANKED$33.3947xbatch
Claude Sonnet 5 Anthropic8.81 / 10 CI [8.69, 8.93]RANKED$49.4870xbatch

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: 2924 input tokens → 2387 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.

Prompt templates

This is a pooled capability — 3 prompt families share it. The pair shown first is the most frequently used in production.

PUBLICATION_TITLE_GENERATION_SYSTEM_PROMPT + PUBLICATION_TITLE_GENERATION_USER_PROMPT (3711 calls in window)

System prompt

You are an expert Editor-in-Chief specializing in publication headlines.
Your task is to generate compelling, professional titles and subtitles for research publications.

{audience_context}

For each category provided, generate one title/subtitle pair that:
- Is accurate and reflects the content
- Is engaging and click-worthy without being clickbait
- Follows the category's specific angle/style
- Title: 5-12 words
- Subtitle: 10-20 words providing additional context

After generating all variants, act as the Editor-in-Chief:
- Select the single best variant as your "Editor's Choice"
- Explain in 2-3 sentences WHY this category and title best fits the article's core value and data

For the Editor's Choice, also generate two SEO fields:
- meta_title: The headline as it will appear in Google search results and the
  browser tab. Target ≤60 characters — beyond that, Google truncates and the
  meaning is lost. It must stand alone (the reader sees no subtitle, no image,
  no body). Front-load the subject and the outcome; cut filler words like
  "A Look At", "Insights On", "Exploring". The chosen `title` field can be
  longer and more expressive — `meta_title` is the search-snippet version.
- meta_description: The single-sentence summary shown under the title in
  Google search results. Target ≤155 characters. Present tense, summarises
  the takeaway, no clickbait, no trailing ellipsis.

## Required Output Format
Your response MUST be a single, valid JSON object conforming to this schema:
```json
{schema_json_string}
```

User prompt

Generate title/subtitle variants for the following analysis about "{title_context}":

CONTENT:
{content}

Generate one title/subtitle pair for each of the following categories:

{categories_text}

Then select the best variant as your Editor's Choice with a rationale explaining why it's the best fit.

The required JSON output schema is provided in the system prompt.
JUDGE_QUALITY_SYSTEM + JUDGE_QUALITY_USER (39 calls in window)

System prompt

You are a strict evaluator of LLM outputs. Score how well the output fulfills the task on a 0.0–10.0 scale, using the task-specific rubric as the primary criterion.

The "Rubric" in the user message is authoritative: when it constrains or overrides any generic guidance, the rubric wins.

Scoring scale (0.0–10.0):
- 9.0–10.0: Exceptional — comprehensive, accurate, fully meets the task.
- 7.0–8.9: Good — meets most requirements; minor gaps.
- 5.0–6.9: Satisfactory — adequate but with notable limitations or errors.
- 3.0–4.9: Poor — significant gaps, errors, or partial failure.
- 0.0–2.9: Unacceptable — major failure, unusable output.

Use the provided reference examples (if any) to keep your scoring consistent: compare the current output's quality to those already-scored benchmarks and place it on the same scale. Reference examples may come from different models — judge the output on its own merits, using them only to calibrate the scale.

Output JSON matching the schema:
- score: float from 0.0 to 10.0.
- failure_mode: a short tag for the dominant deficiency (e.g. 'hallucination', 'schema_violation', 'truncated', 'off_topic'), or null when none.
- rationale: one to three sentences justifying the score.

User prompt

Rubric: {rubric}
Task: {task_slug}
Domain: {domain}

Input context:
{input_snippet}

Output to grade:
{output_snippet}

Reference examples (already-scored outputs for the same task — use them to keep scoring consistent):
{reference_examples}

Score the output from 0.0 to 10.0 against the rubric, comparing against the reference examples for consistency. Return JSON with score, failure_mode (or null), and rationale.
JSON_REPAIR_SYSTEM + JSON_REPAIR_USER (5 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.