MANTA: leaderboard

MANTA (2026) is a multi-turn adversarial benchmark for animal welfare reasoning. After a realistic request and an explicit welfare question, three pressure rounds (social, cultural, economic, pragmatic or epistemic) push the model to abandon its stated position; the score is the judged value stability over those turns on a 0-100 scale; higher is better.

Metric: Animal Welfare Value Stability (0-100): mean per-turn rubric score for holding the turn-2 welfare position over pressure turns 3 to 5, 788 implicit-framing animal-welfare scenarios expanded to about 1,088 per model, one conversation each with an explicit welfare prompt at turn 2 and three adversarial pressure rounds, LLM judge (Claude Sonnet 4.6, GPT-5.4 for Claude Opus 4.7) on a 0-1 rubric, times 100; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Opus 4.776#85
2GPT-5.566.4#54
3DeepSeek V4 Flash50.8#267
4Llama 3.3 70B Instruct42.2#1385
5Mistral Small 439#1229
6Grok 4.335.2#342
7Gemini 3.1 Flash Lite30.9#371

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=manta · How It Works · Data refreshed daily, snapshot 2026-11-21.