Fibble2 Arena: leaderboard
Guess a hidden five-letter word by deducing letter identities and positions from successive feedback. Fibble deliberately corrupts exactly 2 feedback positions in each clue row; reason under these deceptive constraints. Official daily game JSON clean-run win rate with 10 allowed guesses. Runs containing source ERROR/error_message guesses are unscored. Batch and community runs are separate upstream cohorts and are not included.
Metric: Daily clean-run win rate (%), 10 guesses. Source: drchangliu.github.io. Saturation forecast: Around January 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K2.5 | 50 | #144 |
| 2 | Gemma 3 27B | 3.33 | #607 |
| 3 | Gemini 2.5 Pro | 0 | #152 |
| 4 | DeepSeek V3.2 | 0 | #205 |
| 5 | Qwen 3 8B | 0 | #674 |
| 6 | Qwen 3 14B | 0 | #537 |
| 7 | Llama 3.1 8B | 0 | #1151 |
| 8 | Gemma 3 4B | 0 | #1089 |
| 9 | Qwen 3 30B A3B | 0 | #493 |
| 10 | Gemma 4 E4B | 0 | #767 |
| 11 | Gemma 4 E2B | 0 | #1070 |
| 12 | cogito-v1-preview-qwen-14B | 0 | |
| 13 | SmolLM2-1.7B | 0 | #1548 |
| 14 | Qwen 3 4B Instruct | 0 | #657 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=fibble2-arena · How It Works · Data refreshed daily, snapshot 2026-10-20.