Writing and taste
We track 31 benchmarks in writing and taste. They run from everyday level to expert level.
Who held number one in writing and taste
We only count time while other labs are still testing, and stop once a model passes 85%. Faded stretches no longer count.
Anthropic OpenAI
Source: Official leaderboards
Each benchmark counts from the one source its own page leads with: its official leaderboard when that is current, else Epoch AI’s own runs, else labs’ own numbers. We date independent runs by the model’s release and lab reports by when they came out.
Anthropic OpenAI
Source: Official leaderboards
Every independent run: each benchmark’s own leaderboard, Epoch AI’s runs and boards Epoch AI copied, dated by the model’s release. We leave out benchmarks that only labs report.
Anthropic OpenAI
Source: Official leaderboards
Every score, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
From everyday to expert level
Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.
Everyday
Most adults, without training
- DesignArena ASCII Art
- Joke Arena
- Corp Speak Arena
- BuzzBench
- Creation-MMBench
- Slop Score
- Opening Lines Arena
- Stick to Your Role!
- ConStory-Bench
- AidanBench
- Role-play Benchmark (MiniMax)
- LengthCtrl (Tencent)
- Readability (Perplexity)
- Creative Writing (MiniMax in-house)
- DeepSeek Chinese functional writing evaluation
- Slide generation human evaluation (Z.ai)
- Task-dependent output homogenization evaluation
- DeepSeek Chinese creative writing evaluation
University
Undergraduates and entry-level professionals
Expert
Specialists with years of training, often over hours of work
Another 2 have no level. Levels don’t apply to arenas, safety evaluations or composite indices.
By kind
Creative writing
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| Creative Writing v3 Models write responses to 32 creative writing prompts, three times each. A judge model scores them with a rubric and pairwise Elo comparisons. | University | Graded by a model | 83.2 | Dropped |
| Longform Creative Writing (EQ-Bench) Models plan, revise and then write a short story or novella over eight 1,000-word turns from a minimal prompt. A judge model scores quality, and slop, repetition and degradation are tracked. | University | Graded by a model | 86.3 | Not in cards |
| DesignArena ASCII Art Models draw a requested picture using only text characters. Users compare the ASCII art and vote for the better one. | Everyday | Ranked by votes | 1,354 | Not in cards |
| WritingBench Models complete 1,000 real-world writing requests across 6 domains and 100 subdomains. A judge model scores each reply on five query-specific criteria. | University | Graded by a model | 83.9 | Fading |
| Lech Mazur Creative Writing Models write short stories that must use 10 assigned elements such as characters, objects, and motivations. Judge models compare matched stories pairwise. | University | Graded by a model | 3.75 | Not in cards |
| Hemingway-bench Professional writers compare two models' answers blind on real writing prompts, from bedtime stories to business documents. The comparisons become ratings. | Expert | Graded by people | 1,110 | Not in cards |
| Joke Arena Models write jokes. Pairwise votes determine the published Arena rating. | Everyday | Ranked by votes | 1,573 | Not in cards |
| Creation-MMBench Models write creative pieces grounded in up to nine images plus a role, background and instructions, across 765 cases in 51 tasks. A judge scores quality and visual fact alignment against case-specific criteria. | Everyday | Graded by a model |  4.48 | Not in cards |
| Opening Lines Arena Models write opening lines for stories. Pairwise votes determine the published Arena rating. | Everyday | Ranked by votes | 1,613 | Not in cards |
| Stick to Your Role! Models simulate populations of fictional personas and answer personal-values questionnaires in many different conversation contexts. The benchmark measures how stable the expressed values stay across contexts, both for each persona and for the ranking of personas within the population. | Everyday | Checked automatically |  0.90 | Not in cards |
| ConStory-Bench Models write long stories, up to tens of thousands of words, from 2,000 prompts in four task types. An automated checker, ConStory-Checker, finds contradictions in facts, timelines, characters and world rules and cites the passages behind each one. Scores count consistency errors, using a taxonomy of 5 categories and 19 subtypes. | Everyday | Graded by a model | 0.11 | Not in cards |
| AidanBench Models keep giving new answers to open-ended questions until answers become incoherent or repeat earlier ones. The count of fresh, coherent answers is the score, with no ceiling. | Everyday | Graded by a model | 3204.0 | Not in cards |
| Role-play Benchmark (MiniMax) Role-play agents are tested on situated reenactment in Chinese and English scenes. Grading targets the immersive experience rather than a verifiable answer. | Everyday | Graded by a model |  80.6 | Not in cards |
| LengthCtrl (Tencent) Models write text that must meet a requested length, in a Tencent in-house test. Task and scoring details are not published. | Everyday | Graded by a model |  55.4% | Dropped |
| Longform Writing Models write long-form text without tools, capped at 32k output tokens. Moonshot AI has not published the tasks or the scoring method. | University | Graded by a model | 73.8% | Dropped |
| Creative Writing (MiniMax in-house) Models write creative pieces for real Hailuo AI user requests. Automatic assessments score the writing. | Everyday | Graded by a model |  81.3% | System cards only |
| Task-dependent output homogenization evaluation Models answer prompts several times. A judge scores functional diversity where it is wanted and consistency where one answer is right. | Everyday | Graded by a model |  0.21 | System cards only |
| DeepSeek Chinese creative writing evaluation Models write Chinese fiction, prose, poetry and lyrics for 2,837 prompts. Pairwise judging gives separate win rates for instruction following and writing quality. | Everyday | Ranked by matches | System cards only |
Taste and style
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| Judgemark v2.1 The model under test acts as a judge and scores creative writing on several literary criteria. The benchmark measures how well its scores separate stronger writers from weaker ones. | Expert | Checked automatically | 94.1 | Not in cards |
| Judgemark v4 The model under test judges fixed creative-writing samples against low and high anchor responses. The score is how well its ratings separate writer models. | Expert | Checked automatically | 90.7 | Not in cards |
| Corp Speak Arena Models rewrite text in corporate language. Pairwise votes determine the published Arena rating. | Everyday | Ranked by votes | 1,731 | Not in cards |
| BuzzBench Models analyze humor, such as why a joke or comic piece works. The leaderboard reports a score and response length per model. | Everyday | Graded by a model | 71.1 | Not in cards |
| Visual Aesthetic Benchmark Models see a set of images with the same subject and intent and must pick the best one, and in a harder setting both the best and the worst. The 400 sets hold 1,195 commissioned works across fine art, photography and illustration. Labels come from expert judgments, and only sets where experts largely agreed are kept. | Expert | Checked automatically | 40.3% | Not in cards |
| Slop Score Model writing is scored by how often it uses words and patterns that are far more common in AI text than in human text. Lower means less slop. | Everyday | Checked automatically | 18.3 | Not in cards |
| AI Detection Benchmark Systems label private pre-2022 human documents and AI rewrites of them as human or AI written. Both LLMs and commercial detectors are scored by balanced accuracy. | Expert | Checked automatically | 98.4% | Not in cards |
| Design Bench (Moonshot) Models make graphic designs from direct prompts or few-shot reference images. An automated grader rates the results as Elo scores. | University | Graded by a model | 957 | In use |
| Readability (Perplexity) Models answer search questions. Graders rate whether answers are concise yet detailed and use markdown formatting well. | Everyday | Graded by a model |  85.9% | Dropped |
| DeepSeek Chinese functional writing evaluation Models answer 3,170 everyday Chinese writing prompts such as reports, emails and speeches. Pairwise judging against a baseline model gives a win rate. | Everyday | Ranked by matches |  62.6% | System cards only |
| Slide generation human evaluation (Z.ai) People compare HTML slide decks from two models on content, layout and visual quality. Metric is overall win rate. | Everyday | Graded by people | 67.5% | System cards only |
| Slop Arena Site collecting writing-preference data to tune slop detection to a user's taste, paired with the slop-guard tool. | Ranked by votes | Not in cards | ||
| Web UI Bench Many models build the same 12 app UIs from one shared prompt. Outputs are shown side by side for people to compare. No automatic score. | Ranked by votes | Not in cards |