返回
原文
Rohan Paul@rohanpaul_ai
43AI 编辑部评分，满分 100

Grok 4.6 编码成本效率超 GPT-5.6 Sol

2026-08-16 03:17· 14分钟前
AI 导读

Grok 4.6 在三个编码任务中以 $13.11 总成本击败 GPT-5.6 Sol 的 $20.18，模型调用次数 201 次对 338 次，总耗时 129 分 41 秒对 149 分 33 秒。两次运行分别消耗约 20M 和 26M tokens，其中 93-98% 的输入 token 为缓存读取，若无提示缓存成本将增加约 4 倍。

Grok 4.6 beat GPT-5.6 Sol on agentic loop efficiency, spending $13.11 versus $20.18 across the same 3 builds.

Grok 4.6's edge came from bigger coding steps: 201 model calls versus GPT-5.6 Sol's 338.

Really Interesting experiments by @thehypedotnews, a 24/7 AI news in a really nice radio format. (love their chillout music)

1 number in this comparison explains a lot about where coding-agent economics are heading:

93-98% of the input tokens were cache reads.

These runs consumed roughly 20M tokens for Grok 4.6 and 26M for GPT-5.6 Sol, yet the author estimates they would have cost around 4x more without prompt caching.

thehype.grok 4.6 vs gpt 5.6 sol - on three @gameofthrones castles two coding agents built three 3d castles from scratch in a single html file each, then had to render t...
智能体OpenAIxAI编码

来源：Rohan Paul · x.com

导出 Markdown
返回
AI 评分 43/100在 X 查看原推
导出 Markdown

Grok 4.6 编码成本效率超 GPT-5.6 Sol

Rohan Paul · @rohanpaul_ai · X·2026-08-16 03:17·14分钟前
AI 导读

Grok 4.6 在三个编码任务中以 $13.11 总成本击败 GPT-5.6 Sol 的 $20.18，模型调用次数 201 次对 338 次，总耗时 129 分 41 秒对 149 分 33 秒。两次运行分别消耗约 20M 和 26M tokens，其中 93-98% 的输入 token 为缓存读取，若无提示缓存成本将增加约 4 倍。

Grok 4.6 beat GPT-5.6 Sol on agentic loop efficiency, spending $13.11 versus $20.18 across the same 3 builds.

Grok 4.6's edge came from bigger coding steps: 201 model calls versus GPT-5.6 Sol's 338.

Really Interesting experiments by @thehypedotnews, a 24/7 AI news in a really nice radio format. (love their chillout music)

1 number in this comparison explains a lot about where coding-agent economics are heading:

93-98% of the input tokens were cache reads.

These runs consumed roughly 20M tokens for Grok 4.6 and 26M for GPT-5.6 Sol, yet the author estimates they would have cost around 4x more without prompt caching.

thehype.grok 4.6 vs gpt 5.6 sol - on three @gameofthrones castles two coding agents built three 3d castles from scratch in a single html file each, then had to render t...

来源：Rohan Paul· x.com

智能体OpenAIxAI编码评测/基准