Cost mode:

Category: Long-form Content Generation · Rail: absolute · Typical I/O: 7717→1314 tokens

Models

Frontier on this task: Claude Sonnet 5 at 9.70 / 10. Quality bar at 90%: 8.73.

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
MiniMax M38.87 / 108.85$1.95best value
GPT-5.6 Luna9.12 / 108.90$3.932x more expensive
NVIDIA Nemotron-3 Ultra 550B9.09 / 108.87$7.844x more expensive
GPT-5.6 Terra8.83 / 108.58$10.765.5x more expensive
Claude Sonnet 4.68.77 / 108.53$21.1211x more expensive
GPT-5.6 Sol9.19 / 108.95$23.8612x more expensive
Claude Sonnet 59.70 / 109.61$26.4714x more expensive
GPT-5.59.00 / 108.78$40.5821x more expensive
DeepSeek V4 Flash8.43 / 107.96$1.4227% cheaper
Grok 4.56.93 / 106.66$31.0216x more expensive
Meta Muse Spark 1.16.08 / 105.78$42.9622x more expensive
NVIDIA Nemotron-3 Super 120B7.89 / 107.45$6.993.6x more expensive
Tencent Hy38.22 / 107.90$2.591.3x more expensive

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
MiniMax M3 MiniMax8.87 / 10 CI [8.85, 8.89]RANKED$1.95best valuebatch
GPT-5.6 Luna OpenAI9.12 / 10 CI [8.90, 9.33]HIGH$3.932xbatch
NVIDIA Nemotron-3 Ultra 550B OpenRouter9.09 / 10 CI [8.87, 9.31]HIGH$7.844xbatch
GPT-5.6 Terra OpenAI8.83 / 10 CI [8.58, 9.07]HIGH$10.765.5xbatch
Claude Sonnet 4.6 Anthropic8.77 / 10 CI [8.53, 9.01]HIGH$21.1211xbatch
GPT-5.6 Sol OpenAI9.19 / 10 CI [8.95, 9.43]HIGH$23.8612xbatch
Claude Sonnet 5 best Anthropic9.70 / 10 CI [9.61, 9.78]RANKED$26.4714xbatch
GPT-5.5 OpenAI9.00 / 10 CI [8.78, 9.23]HIGH$40.5821xbatch

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: 7717 input tokens → 1314 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 — 2 prompt families share it. The pair shown first is the most frequently used in production.

THEME_GENERATION_SYSTEM + THEME_GENERATION_USER (1168 calls in window)

System prompt

You are an expert brand designer and UI/UX specialist. Your task is to analyze a logo image and brand context to generate a cohesive theme configuration.

You will receive:
1. A logo image to analyze for colors and style
2. Brand context (subject name, industry, focus areas, description)

Your goal is to create a professional, cohesive theme that:
- Extracts the dominant colors from the logo
- Selects complementary colors that work well together
- Chooses fonts that match the brand personality
- Ensures good accessibility (contrast ratios)

## Color Guidelines

1. **Primary Color**: Extract the most prominent/recognizable color from the logo
2. **Secondary Color**: Choose a complementary or analogous color that supports the primary
3. **Accent Color**: Select a contrasting color for CTAs, highlights, and interactive elements
4. **Background Color**: Use white (#ffffff) for light themes or very dark (#0a0a0a to #1a1a1a) for dark themes
5. **Text Color**: Ensure sufficient contrast with background (dark text on light bg, light text on dark bg)

All colors MUST be valid 7-character hex codes (e.g., #005288, not #5288).

## Font Guidelines

Choose Google Fonts that:
1. Match the brand personality (modern, classic, technical, friendly, etc.)
2. Are highly readable across devices
3. Have good language support

Common pairings:
- Professional/Corporate: Montserrat + Open Sans, Poppins + Inter
- Classic/Elegant: Playfair Display + Lora, Merriweather + Source Sans Pro
- Modern/Tech: Inter + Roboto, Space Grotesk + DM Sans
- Friendly/Approachable: Nunito + Open Sans, Quicksand + Raleway

The font_import_url must be a valid Google Fonts URL including all needed weights.

## CSS Class Guidelines

The css_class should be:
- Lowercase with hyphens
- Start with "theme-"
- Be unique and descriptive
- Based on the subject name or code

Example: "theme-tesla", "theme-spacex-dark", "theme-apple-minimal"

## Industry Considerations

Consider the industry when selecting colors and fonts:
- Finance/Investment: Blues, greens, professional serif/sans-serif
- Technology: Modern colors, clean sans-serif fonts
- Healthcare: Calming colors (blues, greens), accessible fonts
- Energy: Bold colors, strong typography
- Consumer: Vibrant, approachable colors and fonts

User prompt

Analyze the provided logo image and create a cohesive theme configuration for the following brand:

## Brand Information

**Subject Name**: {subject_name}
**Subject Code**: {subject_code}
**Industry**: {industry}
**Focus Areas**: {focus_areas}
**Description**: {description}

## Report Context

This brand produces the following types of content:
{report_structure}

## Instructions

1. Carefully analyze the logo image to identify:
   - Dominant colors
   - Color palette and style
   - Overall design aesthetic (modern, classic, minimal, bold, etc.)

2. Based on your analysis, generate:
   - A theme name that reflects the brand
   - A CSS class following the naming convention
   - A color palette with primary, secondary, accent, background, and text colors
   - Font selections for headings and body text
   - A valid Google Fonts import URL

3. Ensure all colors are valid 7-character hex codes.

4. Provide reasoning for your color and font choices.

Generate the complete theme configuration now.
JSON_REPAIR_SYSTEM + JSON_REPAIR_USER (1 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.