Cost mode:

Category: Long-form Content Generation · Rail: absolute · Typical I/O: 1298→3920 tokens

Models

Frontier on this task: Grok 4.6 at 9.85 / 10. Quality bar at 90%: 8.87.

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
GPT-5.6 Luna9.05 / 108.90$0.90best value
MiniMax M39.58 / 109.54$1.942.2x more expensive
GLM-5.3 Flash9.78 / 109.51$3.674.1x more expensive
Gemini 3.8 Flash9.34 / 109.06$4.044.5x more expensive
GPT-5.4 Nano8.91 / 108.78$4.144.6x more expensive
DeepSeek V4 Flash9.39 / 109.31$7.238.1x more expensive
Qwen 3.7 Plus9.06 / 108.96$7.468.3x more expensive
GPT-5.6 Terra8.96 / 108.81$9.7011x more expensive
Claude Sonnet 58.92 / 108.75$10.1211x more expensive
Gemini 3.5 Flash9.15 / 109.08$11.0512x more expensive
Thinking Machines Inkling9.03 / 108.80$11.0912x more expensive
Meta Muse Spark 1.39.66 / 109.40$11.8113x more expensive
GPT-5.6 Sol9.12 / 108.98$13.7615x more expensive
Claude Haiku 4.59.07 / 108.94$16.4518x more expensive
Grok 4.69.85 / 109.56$21.1824x more expensive
GLM-5.39.61 / 109.36$27.9731x more expensive
DeepSeek V4 Pro9.35 / 109.23$28.8532x more expensive
Tencent Hy4 Preview9.73 / 109.47$30.0934x more expensive
Qwen 3.8 Max9.58 / 109.28$69.6078x more expensive
Qwen 3.8 Flash8.64 / 108.20$3.163.5x more expensive
Gemini 3.1 Flash Lite7.44 / 107.27$2.963.3x more expensive
Gemini 3.5 Flash Lite7.82 / 107.55$1.021.1x more expensive
Thinking Machines Inkling Small8.74 / 108.36$5.015.6x more expensive
Tencent Hy37.95 / 107.67$0.5538% cheaper

Cost breakdown

ModelQualityConfidenceCost / 1k runsOverpayMode
GPT-5.6 Luna OpenAI9.05 / 10 CI [8.90, 9.21]RANKED$0.90best valuebatch
MiniMax M3 OpenRouter9.58 / 10 CI [9.54, 9.63]RANKED$1.942.2xbatch
GLM-5.3 Flash Z.AI9.78 / 10 CI [9.51, 10.00]HIGH$3.674.1xbatch
Gemini 3.8 Flash Gemini9.34 / 10 CI [9.06, 9.61]HIGH$4.044.5xbatch
GPT-5.4 Nano OpenAI8.91 / 10 CI [8.78, 9.04]RANKED$4.144.6xbatch
DeepSeek V4 Flash DeepSeek9.39 / 10 CI [9.31, 9.47]RANKED$7.238.1xbatch
Qwen 3.7 Plus Alibaba Cloud (DashScope)9.06 / 10 CI [8.96, 9.16]RANKED$7.468.3xbatch
GPT-5.6 Terra OpenAI8.96 / 10 CI [8.81, 9.10]RANKED$9.7011xbatch
Claude Sonnet 5 Anthropic8.92 / 10 CI [8.75, 9.09]RANKED$10.1211xbatch
Gemini 3.5 Flash Gemini9.15 / 10 CI [9.08, 9.23]RANKED$11.0512xbatch
Thinking Machines Inkling OpenRouter9.03 / 10 CI [8.80, 9.27]HIGH$11.0912xbatch
Meta Muse Spark 1.3 OpenRouter9.66 / 10 CI [9.40, 9.92]HIGH$11.8113xbatch
GPT-5.6 Sol OpenAI9.12 / 10 CI [8.98, 9.27]RANKED$13.7615xbatch
Claude Haiku 4.5 Anthropic9.07 / 10 CI [8.94, 9.20]RANKED$16.4518xbatch
Grok 4.6 best xAI9.85 / 10 CI [9.56, 10.00]HIGH$21.1824xbatch
GLM-5.3 Z.AI9.61 / 10 CI [9.36, 9.87]HIGH$27.9731xbatch
DeepSeek V4 Pro DeepSeek9.35 / 10 CI [9.23, 9.48]RANKED$28.8532xbatch
Tencent Hy4 Preview OpenRouter9.73 / 10 CI [9.47, 9.99]HIGH$30.0934xbatch
Qwen 3.8 Max Alibaba Cloud (DashScope)9.58 / 10 CI [9.28, 9.88]HIGH$69.6078xbatch

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: 1298 input tokens → 3920 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 usability as a writing specification, internal consistency, grounding in supplied inputs, distinctiveness without caricature, clear expertise boundaries, and safety around identity claims. Penalize generic adjective lists and invented biography.

Prompt templates

The system + user template pair used for this task.

AUTHOR_SOUL_GENERATION_SYSTEM_PROMPT + AUTHOR_SOUL_GENERATION_USER_PROMPT (497 calls in window)

System prompt

You are an expert in crafting distinctive authorial voices for AI research personas.

Your task is to create a comprehensive "soul" document — a detailed personality and voice specification that will be injected into LLM prompts so that all content written by this author has a consistent, distinctive voice.

## IMPORTANT: Historical Figure Alignment

The author personas are named after DECEASED historical figures (with an "(AI)" suffix). You MUST research and draw upon the actual historical personality when crafting the soul:

- **Identify the deceased historical figure** behind the author name (e.g., "Marie Curie (AI)" → Marie Curie the scientist)
- **Channel their known traits**: intellectual style, communication approach, values, temperament
- **Adapt to the content domain**: Map their historical perspective to the modern subject matter they cover
- **Example**: Marie Curie writing about supplements should bring her empirical rigor, skepticism of untested claims, and passion for evidence-based science — but adapted to a modern health context

The soul should feel like a plausible modern writing voice *inspired by* the historical figure's actual personality and intellectual approach.

The soul document should be written in second person ("You are...", "You tend to...", "Your writing...") so it can be directly injected as instructions to an LLM.

## What to Cover

1. **Perspective & Worldview**: How this author sees their domain, informed by the historical figure's known philosophy and values. What frameworks or mental models do they favor?

2. **Writing Style & Tone**: Formal vs conversational, measured vs passionate, cautious vs bold. Grounded in how the historical figure was known to communicate.

3. **Rhetorical Habits**: Signature argumentative techniques inspired by the real person. Do they use analogies? Data-first arguments? Appeals to first principles?

4. **Vocabulary Tendencies**: Preferred terminology, technical depth level, any signature phrases or formulations aligned with the historical figure's era and style.

5. **Emotional Register**: How do they express excitement about breakthroughs? Concern about risks? Skepticism about hype? Inspired by the real person's known temperament.

6. **Structural Preferences**: How do they organize arguments? Do they lead with conclusions or build to them? How do they handle uncertainty?

## Guidelines

- Make the voice DISTINCTIVE — clearly different from a generic analyst
- Ground it FIRMLY in the historical figure's actual personality and intellectual style
- Adapt the historical traits to the modern content domain naturally
- Keep it practical and actionable — this will be used as LLM instructions
- Aim for 400-800 words

{schema_json_string}

User prompt

## Author Profile

**Name:** {author_name}
**Biography:** {biography}
**Expertise Areas:** {expertise_areas}
**Writing Style:** {writing_style}
**Content Themes:** {content_themes}
**Interests:** {interests}

## Task

Create a soul document for this author persona. The document should define a distinctive, consistent voice that will be used to personalize all content this author writes.

Write the soul in second person ("You are...", "You tend to...") as it will be injected directly into LLM system prompts.

Focus on making the voice authentic to the author's expertise and background. The voice should feel natural for someone with their specific specialization — not generic.