GPT-6.1 Sol by OpenAI (thinking high) writes the best English today, with a score of 95.3. Among the models you can run on your own machine, Nemotron 3.5 Super VL by NVIDIA (thinking high) leads with 74.3.
The AI Writing Benchmark measures how well leading AI models write in English. Responses to a shared corpus of English writing prompts are compared blind, without model names, and the verdicts are combined into a ranking for English.
115 models compared on 131,698 blind verdicts, updated 12 November 2026.
| # | Model | Score |
|---|---|---|
| 1 |  GPT-6.1 Sol 💡 | 95.3 |
| 2 |  Claude Opus 5.5 💡 | 92.9 |
| 3 |  Kimi K3.1 Pro 💡 | 91.9 |
| 4 | GPT-6 Astra 💡 | 91.5 |
| 5 | Kimi K3 💡 | 90.8 |
| 6 | Claude Fable 5 💡 | 89.0 |
| 7 | Claude Sonnet 5.5 💡 | 88.7 |
| 8 | Claude Fable 5.1 💡 | 88.3 |
| 9 | GPT-6 Sol 💡 | 87.5 |
| 10 | Claude Opus 5 💡 | 87.3 |
This score is not a percentage grade: 50 means an estimated one-in-two chance of beating the average model in English.
Last updated: See the full leaderboard
Ranks recomputed on the duels of one category alone. A model that tops the general ranking does not always top the emails or the fiction. Which model for a novel, an email, a cover letter or an essay: the guide.
| Kind of writing | Model | Score |
|---|---|---|
| Articles, blogs and editorial | GPT-6.1 Sol 💡 | 98.4 |
| Dialogue and screenplay | GPT-6.1 Sol 💡 | 98.5 |
| Explanation and popularization | Kimi K3 💡 | 94.9 |
| Fiction and narrative | Claude Fable 5.1 💡 | 98.9 |
| Marketing and sales | Claude Sonnet 5.5 💡 | 97.9 |
| Personal communication | Claude Opus 5.5 💡 | 98.6 |
| Professional and administrative | Claude Sonnet 5.5 💡 | 93.4 |
| Social media and short formats | GPT-6.1 Sol 💡 | 97.0 |
| Speech and public speaking | GPT-6.1 Sol 💡 | 98.6 |
| Structured documents and reports | GPT-6.1 Sol 💡 | 96.5 |
| # | Model | Score | File | +128K | +256K |
|---|---|---|---|---|---|
| 1 |  Nemotron 3.5 Super VL 💡 | 74.3 | — | — | — |
| 2 |  Muse Glimmer 30B 💡 🏠 | 71.4 | 21.1 GB | 22.9 GB | 22.9 GB |
| 3 | Muse Glimmer 30B 💡 🏠 | 60.2 | 18.2 GB | 20.0 GB | 20.0 GB |
| 4 |  Qwen3.8 27B 💡 🏠 | 57.9 | 17.7 GB | 26.5 GB | 35.1 GB |
| 5 | Qwen3.6 27B 💡 🏠 | 56.9 | 28.6 GB | 37.3 GB | 45.9 GB |
| 6 | Qwen3.8 27B 💡 🏠 | 56.8 | 23.4 GB | 32.1 GB | 40.7 GB |
| 7 | Qwen3.8 27B 💡 🏠 | 55.8 | 30.0 GB | 38.7 GB | 47.3 GB |
| 8 |  Gemma 4 31B 💡 🏠 | 52.4 | 26.4 GB | 38.4 GB | 49.1 GB |
| 9 | Gemma 4 31B 💡 🏠 | 52.4 | 33.8 GB | 45.8 GB | 56.6 GB |
| 10 | Qwen3.6 27B 💡 🏠 | 50.0 | 16.5 GB | 25.3 GB | 33.9 GB |
Size of the GGUF file we tested (GB, decimal, as LM Studio shows it) · +128K: File + context cache at 131,072 tokens (128K), F16 cache, Flash Attention on (GB) · +256K: The same at 262,144 tokens (256K), capped at the native context window (GB)
Best-writing variant that fits the memory of the graphics card or unified memory, short context. Details, speeds and alternatives on the Local AI page.
| Memory | Model | Score |
|---|---|---|
| 8 GB |  Granite 4.2 8B 💡 🏠 | 17.1 |
| 12 GB | Gemma 4 12B 💡 🏠 | 40.0 |
| 16 GB | Gemma 4 12B 💡 🏠 | 40.5 |
| 24 GB | Muse Glimmer 30B 💡 🏠 | 60.2 |
| 32–128 GB | Muse Glimmer 30B 💡 🏠 | 71.4 |
Since this summer, text from Claude and Gemini carries an invisible mark; since 5 October, so does text from ChatGPT and Codex in Europe, and OpenAI's API can be asked to add the mark. Text from Grok, Mistral, Meta AI and models run at home carries none, and Chinese services settle for a visible label. Yet outside the companies involved, nobody can check a text today.
The leaderboard has one winner; your writing does not. On the duels of one kind of text alone, GPT-5.6 Sol takes fiction, Claude Opus 5 personal letters, Kimi K3 formal correspondence, and a local file you can run at home is second for cover letters. A plain guide, task by task.
Same file, same prompts, one switch. Across 21 local builds, turning reasoning on adds 33 points on average for the Qwen family and 10 for Gemma 4 models of 12B and above; the small Gemma E2B and E4B builds gain almost nothing and sometimes lose. Qwen3.8 27B Q4_K_M gains 45 points, more than any other model.
Same model, three files, three scores. Qwen3.8 27B writes English as well at Q4_K_M as at Q8_0 with a file half the size, Gemma 4 31B does not care which of its three builds you run, and Granite 4.2 30B loses twenty points between Q8 and Q4. There is no rule of thumb, only measurements.
The two most-asked-about assistants, judged on the same English prompts in blind duels: which one writes better, by how much, and for which kinds of text. Updated with every export of the ranking.
OpenAI's best model against Google's best, on the same English prompts, in blind duels: who leads on the leaderboard, how the direct duels went, and where each one is stronger.
Anthropic's best model against Google's best, on the same English prompts, in blind duels: the gap on the leaderboard, the direct record, and the kinds of writing where Gemini comes closest.
All head-to-heads and the free comparison →
One prompt, two anonymous texts: pick the better one and feed the human ranking. The model names appear after your choice.
Choose any two models: their scores, their direct duels and a published pair of texts side by side, by kind of writing.
How each AI judge ranks the models, with wins, ties and losses per judge, and where the judges disagree.
Measured tokens per second, memory footprint and writing score of every local file on a Ryzen AI Max+ 395 with 128 GB.
Who runs this benchmark, on which machines, and the principles behind the blind duels.
The ranking evaluates writing quality alone. Prompts are written directly in the language being tested, and model names remain hidden until the verdict so reputation cannot influence the evaluation. This site is a personal project. No model provider pays me, and I have no reason to favour one model over another.
