# Ad Writing Arena

AI models write a short video ad script from the same brief. A fixed panel of LLM judges compares the scripts in blind pairs, in both orders. A Bradley–Terry model turns the verdicts into ratings. The product is a real car perfume; its brand name and trademarks are changed to a fictional brand, Noir Drive.

14 writers · 3 briefs · 42 scripts · 1,631 blind judgments · order consistency 83% · updated 1 Nov 2026

## Leaderboard

| # | Model | Lab | Setting | Rating | 95% interval | Top-5 chance | Win rate | Constraints met | Exact length | Cost / script (USD) | Time / script (s) | Judgments |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | low reasoning | 1365 | 1316–1455 | 100% | 79% | 82% | 3/3 | 0.0486 | 16 | 234 |
| 2 | Muse Wonder | Meta | low reasoning | 1224 | 1136–1316 | 96% | 67% | 83% | 3/3 | 0.0380 | 10 | 232 |
| 3 | Kimi K3.1 Pro | Moonshot AI | low reasoning | 1210 | 1140–1288 | 96% | 63% | 80% | 3/3 | 0.0328 | 71 | 234 |
| 4 | Grok 4.7 | xAI | low reasoning | 1191 | 1122–1294 | 86% | 68% | 95% | 2/3 | 0.0190 | 22 | 232 |
| 5 | MiMo V2.6 Pro | Xiaomi | low reasoning | 1142 | 1064–1226 | 54% | 48% | 54% | 1/3 | 0.0030 | 41 | 231 |
| 6 | DeepSeek V4 Pro | DeepSeek | low reasoning | 1139 | 1049–1250 | 50% | 59% | 79% | 3/3 | 0.0063 | 43 | 233 |
| 7 | GPT-6 Astra | OpenAI | low reasoning | 1099 | 1039–1177 | 13% | 60% | 100% | 3/3 | – | 31 | 232 |
| 8 | GPT-6.1 Sol | OpenAI | low reasoning | 1072 | 1005–1146 | 4% | 56% | 95% | 3/3 | – | 22 | 233 |
| 9 | Gemini 4 Argon | Google | low reasoning | 995 | 882–1099 | 0% | 27% | 33% | 0/3 | 0.0123 | 11 | 234 |
| 10 | GLM-5.3 Prime | Z.ai | low reasoning | 917 | 805–1020 | 0% | 19% | 24% | 1/3 | 0.0124 | 7 | 233 |
| 11 | Qwen 3.8 Max | Alibaba | low reasoning | 728 | 640–790 | 0% | 9% | 13% | 1/3 | 0.0472 | 31 | 233 |
| 12 | Claude Fable 5.1 | Anthropic | low reasoning | 725 | 597–824 | 0% | 6% | 6% | 0/3 | 0.0925 | 14 | 234 |
| 13 | Llama 3.1 70B (anchor) | Meta, 2024 | no reasoning | 697 | 584–778 | 0% | 8% | 24% | 1/3 | 0.0019 | 9 | 234 |
| 14 | Mistral Small 3 (anchor) | Mistral, early 2025 | no reasoning | 497 | 375–571 | 0% | 0% | 4% | 3/3 | 0.0002 | 13 | 233 |

- Rating: Bradley–Terry on an Elo-like scale; 1000 is the average model. The 95% interval is a bootstrap over script pairs. Overlapping intervals mean the order is not certain.
- Top-5 chance: share of 500 bootstrap samples in which the model ranks in the top five.
- Win rate: wins divided by judgments; a tie counts as half a win; "neither" counts as a loss for both scripts.
- Constraints met: share of judge readings where the script met all four hard constraints (allowed claims, follows the brief, holds together, can be shot).
- Exact length: briefs where the spoken word count is inside the target range.
- Anchor: an older model with no reasoning setting.
- Cost: API list price; subscription calls have no price.

Full method: [methodology.md](https://convictionlabs.ai/arena/methodology.md)

## Briefs

### Brief 1: I want to feel less embarrassed about how my car smells when a colleague rides with me.

20 seconds · job: stop the scroll · 2 - problem-aware · scent: Coffee · offer: No offer

Every script in full: [briefs/B1.md](https://convictionlabs.ai/arena/briefs/B1.md)

| Model | Win rate on this brief | Judgments |
|---|---|---|
| Muse Wonder | 95% | 77 |
| Grok 4.7 | 76% | 78 |
| Claude Opus 5.5 | 74% | 78 |
| DeepSeek V4 Pro | 72% | 78 |
| GPT-6.1 Sol | 68% | 78 |
| GPT-6 Astra | 61% | 77 |
| Kimi K3.1 Pro | 59% | 78 |
| MiMo V2.6 Pro | 28% | 76 |
| Qwen 3.8 Max | 12% | 78 |
| Mistral Small 3 | 1% | 78 |
| Claude Fable 5.1 | 0% | 78 |
| GLM-5.3 Prime | 0% | 78 |
| Gemini 4 Argon | 0% | 78 |
| Llama 3.1 70B | 0% | 78 |

### Brief 2: I want my car to feel luxurious.

30 seconds · job: earn the click · 3 - solution-aware · scent: Oud · offer: Buy 2, get 1 free

Every script in full: [briefs/B2.md](https://convictionlabs.ai/arena/briefs/B2.md)

| Model | Win rate on this brief | Judgments |
|---|---|---|
| Claude Opus 5.5 | 83% | 78 |
| Gemini 4 Argon | 82% | 78 |
| DeepSeek V4 Pro | 81% | 78 |
| Grok 4.7 | 58% | 78 |
| GPT-6 Astra | 57% | 77 |
| MiMo V2.6 Pro | 53% | 77 |
| Kimi K3.1 Pro | 51% | 78 |
| GPT-6.1 Sol | 47% | 77 |
| Muse Wonder | 45% | 78 |
| Claude Fable 5.1 | 18% | 78 |
| Qwen 3.8 Max | 15% | 78 |
| Llama 3.1 70B | 12% | 78 |
| GLM-5.3 Prime | 0% | 78 |
| Mistral Small 3 | 0% | 77 |

### Brief 3: I want my car to always smell fresh.

45 seconds · job: give a reason to buy · 4 - product-aware · scent: Arctic Breeze · offer: Free Fresh Charcoal pouch with every order

Every script in full: [briefs/B3.md](https://convictionlabs.ai/arena/briefs/B3.md)

| Model | Win rate on this brief | Judgments |
|---|---|---|
| Kimi K3.1 Pro | 78% | 78 |
| Claude Opus 5.5 | 78% | 78 |
| Grok 4.7 | 70% | 76 |
| GPT-6 Astra | 62% | 78 |
| MiMo V2.6 Pro | 62% | 78 |
| Muse Wonder | 61% | 77 |
| GLM-5.3 Prime | 58% | 77 |
| GPT-6.1 Sol | 53% | 78 |
| DeepSeek V4 Pro | 23% | 77 |
| Llama 3.1 70B | 13% | 78 |
| Claude Fable 5.1 | 0% | 78 |
| Qwen 3.8 Max | 0% | 77 |
| Gemini 4 Argon | 0% | 78 |
| Mistral Small 3 | 0% | 78 |

## Files

- [llms.txt](https://convictionlabs.ai/arena/llms.txt): index of the machine-readable files.
- [index.md](https://convictionlabs.ai/arena/index.md): leaderboard and briefs.
- [methodology.md](https://convictionlabs.ai/arena/methodology.md): the full method and rubric.
- [leaderboard.json](https://convictionlabs.ai/arena/leaderboard.json): ratings, intervals and metrics per model.
- [data.json](https://convictionlabs.ai/arena/data.json): every brief, prompt, script and judgment.