Testing AI on taste.
We benchmark frontier models on the judgments working creatives make every day: which cover, which crop, which line of copy. Every item is judged blind by a panel of professionals, and we build every benchmark in-house.
- 72.4Claude Opus 5.5
- 70.1GPT-6 Astra
- 64.8Gemini 3.8 Flash
- 60.2Muse Spark 1.3
- 58.9Grok 4.7
- 56.4Kimi K3
- 54.7Qwen3.8-Max
- 51.9DeepSeek-V4.1-Flash
- 49.3GLM-5.3
- 44.6Mistral Medium 3.5
Latest reports
Recent benchmark releases and model evaluations.
Anthropic’s Claude Opus 5.5 evaluated across our benchmark suite
| Benchmark | Bar | Accuracy | Rank |
|---|---|---|---|
| FrontierTaste Benchmark | 78.0% ±1.60 | 1/12 | |
| Cover Pairs v1.1 | 73.2% ±2.30 | 1/12 | |
| Type Specimen Bench | 72.2% ±3.00 | 1/12 | |
| Photo Crop Judge | 82.8% ±3.70 | 1/11 | |
| Layout Critique | 71.4% ±1.60 | 1/12 | |
| Motion Timing | 61.8% ±2.30 | 3/10 | |
| Copy Voice | 82.0% ±3.00 | 1/12 | |
| Brand Consistency | 76.2% ±3.70 | 1/9 | |
| AestheticQA | 89.2% ±1.60 | 1/26 |
We evaluated Anthropic’s new Claude Opus 5.5 across the FrontierTaste suite. In this preview it leads the FrontierTaste Benchmark and Layout Critique.
Motion remains its weakest discipline: in most disagreements with the panel it prefers the busier animation, and its written reasons rarely mention timing.
Five working professionals judge every pair, blind. A model scores only when it agrees with them.
Discipline leaderboard
Model taste across the crafts that make up creative work.
Cover Pairs v1.1
Can models tell the better book cover?
Full details soon- Claude Opus 5.5
- GPT-6 Astra
- Claude Fable 5.1
- Gemini 3.8 Flash
- GPT-6 Sol
- Muse Spark 1.3
- Grok 4.7
- Kimi K3
- Qwen3.8-Max
- DeepSeek-V4.1-Flash
- GLM-5.3
- Mistral Medium 3.5
Model taste over time
How each lab’s best FrontierTaste Benchmark score has moved with each release.
- GPT-5.1, Nov ’25: 41.6
- GPT-5.6, Jul ’26: 60.3
- GPT-6 Astra, Sep ’26: 70.1
- Gemini 3 Pro, Nov ’25: 43.0
- Gemini 3.6 Flash, Jul ’26: 57.2
- Gemini 3.8 Flash, Sep ’26: 64.8
- Muse Spark 1.3, Sep ’26: 60.2
- Grok 4.1, Nov ’25: 36.8
- Grok 4.5, Jul ’26: 51.4
- Grok 4.6, Aug ’26: 55.1
- Grok 4.7, Sep ’26: 58.9
- Kimi K2 Thinking, Nov ’25: 38.9
- Kimi K3, Jul ’26: 56.4
- Qwen3.8-Max, Aug ’26: 54.7
- DeepSeek-V3.2, Dec ’25: 35.2
- DeepSeek-V4.1-Flash, Sep ’26: 51.9
- GLM-4.6, Oct ’25: 30.4
- GLM-5.3, Aug ’26: 49.3
- Mistral Medium 3.5, Apr ’26: 44.6
- Claude Opus 4.5, Nov ’25: 44.2
- Claude Opus 5, Jul ’26: 63.8
- Claude Opus 5.5, Sep ’26: 72.4
Preview: model names are real; scores, costs and timings on this page are illustrative placeholders, not published results.
