ani-cli AI index / bench-based language model suggestions

[port19x]

Here is how I'd bench it, frontiercode for code quality, terminalbench because of high shell scripting correlation and the two security benches due to our relationship with the anime providers.

Formula

50% FrontierCode 1.1
25% Terminal-Bench 4.0
15% DeepsecBench-AA
10% CWE-Bench-AA

Each one normalised against the top score.
Running this today gives me the following leaderboard.

Table for Agents

# | Model | Index | FrontierCode | TB 4.0 | Deepsec | CWE
-- | -- | -- | -- | -- | -- | --
1 | Kimi K3.1 Pro | 90.8 | 90 | 100 | 79 | 88
2 | Claude Opus 5.5 | 89.3 | 100 | 89 | 58 | 84
3 | GPT-6 Astra | 88.1 | 90 | 88 | 80 | 91
4 | Claude Fable 5.1 | 81.1 | 94 | 77 | 50 | 74
5 | MiMo-V2.6-Pro | 74.0 | 87 | 52 | 56 | 91
6 | DeepSeek V4.1 Flash 1017 | 72.5 | 84 | 50 | 54 | 100
7 | GLM-5.3 | 69.9 | 77 | 62 | 52 | 81
8 | GLM-5.3-Flash | 66.0 | 79 | 49 | 43 | 81
9 | Mistral Large 4 Preview | 62.3 | 80 | 40 | 35 | 74
10 | Step 5 Preview | 62.0 | 74 | 50 | 35 | 77
11 | GPT-6 Luna | 60.8 | 80 | 19 | 52 | 82
12 | Grok 4.7 | 60.6 | 65 | 38 | 58 | 99
13 | Gemini 3.8 Flash | 54.7 | 70 | 29 | 32 | 77
14 | Qwen3.8 27B | 42.6 | 66 | 8 | 26 | 39
15 | Beam 1.0 | 33.7 | 47 | 4 | 15 | 71

[justchokingaround]

lol, now this is interesting

[port19x]

ofc AA has to release the intelligence index V5 less than 24 hours later 🙄