gpt-6IMAGE

Test: Astra and other models' output speed cut to 60%

The author measured tok/s by dividing total output tokens by total time over 10 runs, finding Astra, 5.6 Sol and Luna dropped to 60% of their previous speed and GPT-6.1 Sol to just 15 tok/s, suspecting OpenAI traded speed for quota.

lifcc

@mylifccArchived 2026.10.02Originally posted 2026.10.01

Test: Astra and other models' output speed cut to 60%

lifcc @mylifccOriginally posted 2026.10.01

Open original on X

OpenAI为了解决大家额度不够的问题，把所有模型的速度降低了一半，以前你能用1一天，现在能用1天半了，我哭死通过10 次的总输出 token ÷ 10 次的总耗时，我拿到了每个模型的tok/s，而且这个是包含思考的（没有内容只有数量统计）。 Astra、5.6Sol、Luna的速度全部被降低到了原来的60% GPT-6.1 sol 更是降低到15tok/s

x.com/mylifcc/status/2105502252345835608Archived 2026.10.02 10:04