Can it write like me?
Turn neutral drafts of the site author's real messages back into the author's own voice.
Run Nov 4, 2026, 53 items, $62.20 to run.
Method
The site's author chose 53 of their longest English messages, 74 to 211 words each: chat replies, Discord posts and prompts typed to AI assistants between 2023 and 2026, every one reviewed by the author as entirely their own words. A separate model rewrote each one in plain, neutral English with the same content, and a second check dropped any draft that changed the meaning.
Each model sees 30 other messages by the author as samples of the style, then one neutral draft at a time, and is asked to rewrite it so that it reads as if the author wrote it, keeping the meaning. One request per model and message, at the highest reasoning effort each model supports, with a 32,768-token output cap including reasoning and no tools.
A classifier trained on the author's other English writing, against other people's ChatGPT prompts and chat messages and against GPT- and Claude-written text, gives the probability that the author wrote each rewrite. It sees function words, punctuation, capitalisation and word shapes, not topics. A rewrite that changes the meaning, answers the message or refuses scores 0. The score is the mean probability as a percentage; only models answering at least 90% of the messages are ranked.
The messages, the drafts and every rewrite are private; this page publishes the method and the scores.
Results
| # | Model | Style match |
|---|---|---|
| The author's own messages | 79.0% | |
| 1 | Claude Fable 5.1 (max reasoning · 30 samples) | 76.7% |
| 2 | GPT-6 Astra Pro (max reasoning · 30 samples) | 75.6% |
| 3 | GPT-6.1 Pro Sol (max reasoning · 30 samples) | 75.4% |
| 4 | GPT-6 Astra (max reasoning · 30 samples) | 75.1% |
| 5 | Claude Opus 5.5 (max reasoning · 30 samples) | 73.7% |
| 6 | GPT-6.1 Sol (max reasoning · 30 samples) | 72.5% |
| 7 | GPT-6.1 Terra (max reasoning · 30 samples) | 68.1% |
| 8 | Claude Sonnet 5.5 (max reasoning · 30 samples) | 67.4% |
| 9 | Gemini 4 Argon (high reasoning · 30 samples) | 66.2% |
| 10 | Muse Wonder (max reasoning · 30 samples) | 64.9% |
| 11 | Kimi K3.1 Pro (max reasoning · 30 samples) | 63.1% |
| 12 | Grok 4.7 (xhigh reasoning · 30 samples) | 57.5% |
| 13 | DeepSeek V4.1 Flash (1017) (max reasoning · 30 samples) | 53.2% |
| The neutral drafts, unchanged | 46.0% |
- Two reference rows bound the scale: the author's own held-out messages, scored by the same classifier, and the neutral drafts left unchanged. A model that fully took on the author's voice would score near the first; one that only tidied the draft, near the second.
- One author's informal English: the board measures imitation of one person from 30 samples, not writing quality or style in general. Close scores do not establish an ordering.
In the main leaderboard
These results are collected for the main leaderboard alongside external benchmarks. New scores enter the rating through the regular audited data updates. Comparisons with that rating therefore share some of its inputs.
Interactive version: theaggregate.ai/our-benchmarks/write-like-me · How It Works · Data refreshed daily, snapshot 2026-11-05.