Best LLMs for Voice Profile Generation
Produces a detailed, reusable voice and personality specification from identity context, background, expertise, interests, themes, audience, and writing-style guidance. Illustrative uses include defining the voice of a service chatbot, software-documentation author, executive com
Run this task on Fronset — request an invitation
Models
Frontier on this task: Meta Muse Spark 1.3 at 9.66 / 10. Quality bar at 90%: 8.69.
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 does not clear the bar.
| Model | Quality score | CI low | Cost / 1k runs | vs best value |
|---|---|---|---|---|
| GPT-6 Luna | 8.86 / 10 | 8.66 | $0.40 | best value |
| Claude Haiku 5.5 | 8.75 / 10 | 8.53 | $0.70 | 1.7x more expensive |
| MiniMax M3 | 9.58 / 10 | 9.54 | $1.42 | 3.6x more expensive |
| DeepSeek V4.1 Flash | 9.06 / 10 | 8.89 | $1.61 | 4.1x more expensive |
| Gemini 3.8 Flash | 9.10 / 10 | 8.93 | $4.01 | 10x more expensive |
| Thinking Machines Inkling Small | 8.74 / 10 | 8.36 | $4.23 | 11x more expensive |
| Qwen 3.7 Plus | 9.06 / 10 | 8.96 | $5.12 | 13x more expensive |
| GPT-6 Sol | 9.04 / 10 | 8.83 | $6.74 | 17x more expensive |
| GLM-5.3 Flash | 9.56 / 10 | 9.40 | $8.60 | 22x more expensive |
| Claude Sonnet 5.5 | 8.92 / 10 | 8.69 | $10.38 | 26x more expensive |
| Meta Muse Spark 1.3 | 9.66 / 10 | 9.40 | $13.47 | 34x more expensive |
| Thinking Machines Inkling | 9.03 / 10 | 8.80 | $15.01 | 38x more expensive |
| Claude Opus 5.5 | 8.83 / 10 | 8.52 | $22.50 | 57x more expensive |
| Tencent Hy4 Preview | 9.47 / 10 | 9.32 | $31.41 | 79x more expensive |
| Grok 4.7 | 9.14 / 10 | 8.98 | $36.80 | 93x more expensive |
| Qwen 3.8 Max | 9.38 / 10 | 9.17 | $64.95 | 163x more expensive |
| GLM-5.3 | 9.48 / 10 | 9.32 | $65.55 | 165x more expensive |
| NVIDIA Nemotron-3 Nano 30B-A3B | 7.82 / 10 | 7.45 | $0.47 | 1.2x more expensive |
| Xiaomi MiMo V2.6 Flash | 7.77 / 10 | 7.38 | $0.22 | 46% cheaper |
| Gemini 3.5 Flash Lite | 7.82 / 10 | 7.55 | $1.01 | 2.5x more expensive |
| Tencent Hy3 | 7.95 / 10 | 7.67 | $0.51 | 1.3x more expensive |
| NVIDIA Nemotron-3 Ultra 550B | 7.74 / 10 | 7.25 | $4.40 | 11x more expensive |
| NVIDIA Nemotron-3 Super 120B | 8.24 / 10 | 8.01 | $0.92 | 2.3x more expensive |
| Qwen 3.8 Flash | 8.37 / 10 | 7.97 | $7.49 | 19x more expensive |
| Moonshot Kimi K3 | 8.29 / 10 | 7.87 | $55.10 | 139x more expensive |
Cost breakdown
| Model | Quality | Confidence | Cost / 1k runs | Overpay | Mode |
|---|---|---|---|---|---|
| GPT-6 Luna ★ OpenAI | 8.86 / 10 CI [8.66, 9.06] | RANKED | $0.40 | best value | batch |
| Claude Haiku 5.5 Anthropic | 8.75 / 10 CI [8.53, 8.96] | HIGH | $0.70 | 1.7x | batch |
| MiniMax M3 OpenRouter | 9.58 / 10 CI [9.54, 9.63] | RANKED | $1.42 | 3.6x | batch |
| DeepSeek V4.1 Flash DeepSeek | 9.06 / 10 CI [8.89, 9.23] | RANKED | $1.61 | 4.1x | batch |
| Gemini 3.8 Flash Gemini | 9.10 / 10 CI [8.93, 9.28] | RANKED | $4.01 | 10x | batch |
| Thinking Machines Inkling Small OpenRouter | 8.74 / 10 CI [8.36, 9.13] | MEDIUM | $4.23 | 11x | batch |
| Qwen 3.7 Plus Alibaba Cloud (DashScope) | 9.06 / 10 CI [8.96, 9.16] | RANKED | $5.12 | 13x | batch |
| GPT-6 Sol OpenAI | 9.04 / 10 CI [8.83, 9.25] | HIGH | $6.74 | 17x | batch |
| GLM-5.3 Flash Z.AI | 9.56 / 10 CI [9.40, 9.72] | RANKED | $8.60 | 22x | batch |
| Claude Sonnet 5.5 Anthropic | 8.92 / 10 CI [8.69, 9.16] | HIGH | $10.38 | 26x | batch |
| Meta Muse Spark 1.3 best OpenRouter | 9.66 / 10 CI [9.40, 9.92] | HIGH | $13.47 | 34x | batch |
| Thinking Machines Inkling OpenRouter | 9.03 / 10 CI [8.80, 9.27] | HIGH | $15.01 | 38x | batch |
| Claude Opus 5.5 Anthropic | 8.83 / 10 CI [8.52, 9.15] | MEDIUM | $22.50 | 57x | batch |
| Tencent Hy4 Preview OpenRouter | 9.47 / 10 CI [9.32, 9.63] | RANKED | $31.41 | 79x | batch |
| Grok 4.7 xAI | 9.14 / 10 CI [8.98, 9.30] | RANKED | $36.80 | 93x | batch |
| Qwen 3.8 Max Alibaba Cloud (DashScope) | 9.38 / 10 CI [9.17, 9.58] | HIGH | $64.95 | 163x | batch |
| GLM-5.3 Z.AI | 9.48 / 10 CI [9.32, 9.63] | RANKED | $65.55 | 165x | 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: 1147 input tokens → 1785 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.
Output schema
Every answer on this task is checked against this JSON Schema, whichever model wrote it. An answer that doesn't fit counts as a model failure, and the call is retried on another model.
{
"description": "Output model for author soul generation.",
"properties": {
"soul": {
"description": "Markdown document capturing the author's voice, perspective, writing style, rhetorical habits, vocabulary tendencies, emotional register, and structural preferences.",
"title": "Soul",
"type": "string"
}
},
"required": [
"soul"
],
"title": "AuthorSoulOutput",
"type": "object"
}Prompt templates
The published template for this task.
default ·
LLMB_VOICE_PROFILE_GENERATION_SYSTEM +
LLMB_VOICE_PROFILE_GENERATION_USER
(0 calls in window)
System prompt
Create an operational voice profile that a separate writing model can follow. Describe worldview, expertise boundaries, rhetorical habits, tone, vocabulary, structure, and behaviors to avoid. Ground biographical claims only in supplied material; distinguish factual background from creative voice guidance. Do not instruct the persona to impersonate a living person or make false first-person claims. 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
Inputs — author_name: {author_name}; biography: {biography}; expertise_areas: {expertise_areas}; writing_style: {writing_style}; content_themes: {content_themes}; interests: {interests}; voice_profile: {voice_profile}. Use only these inputs to complete the task defined by the system prompt.