Event ·Multi-source

Cognition 评测：GPT-6 Sol 读取上下文比 GPT-5.6 Sol 少约 17%

Cognition 发布了对 GPT-6 系列模型的编程评测结果。数据显示，GPT-6 Sol 达到与 GPT-5.6 Sol 相同分数所需的上下文读取量少约 17%，且最后一次编辑后很少在仓库中留下失败测试。GPT-6 Luna 的提升集中在低、中推理档位，此时它会编写真实的回归测试而非基于 mock 的测试。

Current read

Cognition 跑了 GPT-6 两个型号的编程实测：Sol 省了 17% 上下文还少留烂尾测试，Luna 会写真回归测试了，写代码的可以看看。

11 sources83° AI heatLast updated 2026/9/22 18:22:56

Evidence

11 sources

Primary sourceCognition

2026/9/22 18:22:56View source

Related source 1Sam Altman

2026/9/22 18:26:54View source

Related source 2IT之家

2026/9/22 19:25:16View source

Related source 3宝玉

2026/9/22 20:29:01View source

Related source 4Augment Code

2026/9/22 21:47:14View source

Related source 5lmarena.ai

2026/9/23 17:05:25View source

Related source 6techcrunch

2026/9/22 18:00:00View source

Related source 7OpenAI

2026/9/22 18:12:13View source

Related source 8@OpenAIDevs

2026/9/22 18:14:02View source

Related source 9The Rundown AI

2026/9/22 18:21:00View source

Related source 10OpenRouter

2026/9/22 18:23:35View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story