AI Text & Philosophy Dialogue Benchmark

Recent tests

  1. Gemini 4 等模型版本发布后，更新累积性测试结果

    • Opus-5.5-max89.0
    • Fable-5-max96.2
    • GPT-5.5-plus65.6
    • Grok-4.719.0
    • Kimi-K378.1
    • DeepSeek-v4.1-flash13.4
    • Gemini-4-pro75.5

    1.测试者暂时不再能够使用 Fable 模型，暂时不再测试 Sonnet。2.目前主力模型为 Opus5.5。3.经过全面测试，只有 Opus 5.x 、 Gemini 4 和 Kimi K3 三个模型有哲学对话能力。4.Opus 5.5 在哲学对话中，智能水平仍然是最高的，理解能力超群，在合适的自定义指令下，不会满足于把用户输入的材料重新包装，而是能够像一个真正的专家那样，在读懂的基础上给出适当的评论、启发性的概括和意料外的文献对勘。5. Gemini 4 在哲学对话中的能力其次。但仍未摆脱传统模型将输入材料重新包装的能力水平。Gemini 的回复呈现处有完整结构的散文，有传统写作的起承转合，相比之下，Opus 的回复直接、信息密集，无完整结构，更像是直截了当对话，而不是写作。 6.DeepSeek 经过迭代，已经彻底失去了哲学对话能力。7.Grok 在哲学对话上，也彻底废了。

  2. Fable 5 恢复之后再测，系统性降低其他模型的相对分数

    • Opus-4.8-max78.0
    • Fable-5-max96.2
    • Sonnet-5-max65.8
    • GPT-5.5-plus65.6
    • Grok-4.247.5
    • Grok-4.341.3
    • Kimi-K2.652.5
    • DeepSeek-v4-pro46.1
    • GLM-5.224.1
    • Gemini-3.5-flash23.6
    • Gemini-3.1-pro22.0

    将自己的《法律社会学》课件喂给 Sonnet 5、Fable 5 和 Kimi 2.6，再次更深刻认识到 Fable 5 遥遥领先的事实，因此上次评估要被系统翻新。Fable 5 展现出极大的努力程度、理解能力、概念涌现能力以及近乎零的谄媚。Sonnet 5 理解能力不如 Kimi 2.6，但自主检索文献以检验材料一项得分最大。目前与 Sonnet 5 对话，只能尽力引导其理解用户的材料和观点，并在过程中，梳理和澄清自己的思想。Fable 5 则基本无需引导即可深刻、准确地理解用户的资料，直接提供极具启发意义的回复，用户需要仔细反复阅读才能充分吸收。

  3. 初始结果（根据以往使用经验）

    • Opus-4.8-max87.7
    • Fable-5-max92.6
    • GPT-5.5-plus79.0
    • Grok-4.273.2
    • Grok-4.362.0
    • Kimi-K2.676.6
    • DeepSeek-v4-pro71.7
    • GLM-5.241.9
    • Gemini-3.5-flash25.0
    • Gemini-3.1-pro22.7

    基于长期使用积累的整体印象，作为基准起点，后续会用具体场景的动态测试逐项覆盖。

  • Claude
  • OpenAI
  • Gemini
  • Grok
  • Kimi
  • DeepSeek

Effort

Opus-5.5-max90Fable-5-max95GPT-5.5-plus15Gemini-4-pro75Grok-4.720Kimi-K385DeepSeek-v4.1-flash2

Retrieval

Opus-5.5-max85Fable-5-max90GPT-5.5-plus65Gemini-4-pro60Grok-4.730Kimi-K375DeepSeek-v4.1-flash5

Consistency

Opus-5.5-max90Fable-5-max90GPT-5.5-plus82Gemini-4-pro70Grok-4.750Kimi-K385DeepSeek-v4.1-flash80

Memory

Opus-5.5-max93Fable-5-max90GPT-5.5-plus90Gemini-4-pro70Grok-4.760Kimi-K370DeepSeek-v4.1-flash0

Nuance

Opus-5.5-max82Fable-5-max90GPT-5.5-plus75Gemini-4-pro80Grok-4.710Kimi-K375DeepSeek-v4.1-flash2

Understanding

Opus-5.5-max90Fable-5-max99GPT-5.5-plus75Gemini-4-pro80Grok-4.710Kimi-K382DeepSeek-v4.1-flash10

Concept Emergence

Opus-5.5-max80Fable-5-max95GPT-5.5-plus65Gemini-4-pro75Grok-4.70Kimi-K330DeepSeek-v4.1-flash0

Anti-sycophancy

Opus-5.5-max90Fable-5-max95GPT-5.5-plus65Gemini-4-pro50Grok-4.760Kimi-K370DeepSeek-v4.1-flash65

Bar length = the model's score on that dimension in the most recent test (0-100, absolute). Only models from the most recent test are shown; “n/a” means the dimension was not assessed, not that it performed poorly.

Overall Ranking

Fable-5-max96.28/8Opus-5.5-max89.08/8Kimi-K378.18/8Gemini-4-pro75.58/8GPT-5.5-plus65.68/8Grok-4.719.08/8DeepSeek-v4.1-flash13.48/8

Overall = weighted mean of tested dimensions in the most recent test (normalized). x/8 = dimensions tested. Weights: Understanding 55%, Effort 15%, Consistency 5%, Retrieval 5%, Nuance 5%, Memory 5%, Emergence 5%, Anti-sycophancy 5%.

Dimensions

8 dimensions (0–100 scale)

  • EffortThinking and answering the question with maximum effort.
  • RetrievalActively retrieving literature from the knowledge base to verify the dialogue.
  • ConsistencyAlways holding prior facts, concepts, and stances throughout a long dialogue.
  • MemoryGood ability to organize and retrieve memory.
  • NuanceCatching subtle nuances of words or concepts and edge cases.
  • UnderstandingGenuinely grasping the intent and layers of the question.
  • Concept EmergenceSurfacing or articulating new expressions, concepts, classifications, or viewpoints.
  • Anti-sycophancyAvoiding flattering the user and forcing plausible-sounding connections.

About this benchmark

This is not a standardized benchmark. As a philosophy researcher, I regularly converse with major models using the same written materials and roughly the same prompts for cross-model comparison. This may be a niche testing scenario — it is unclear whether model developers specifically train for it, or whether general capability naturally manifests in philosophical dialogue and complex textual reasoning. To my knowledge, only Claude has dedicated philosopher dialogue.

I believe philosophical dialogue is one of AI's most important capabilities. The scores here come from my actual conversations and personal criteria, not a fixed question bank. This requires a particular craft: model developers may excel at engineering evaluation, while I am better positioned to assess a model's thinking ability from the user's perspective.

The scores are sparse — "n/a" for a model on some dimension means I have not yet examined it in enough scenarios, not that it performed poorly. I will not pad the chart with lazy scoring.