4 versions of glm, same 3 prompts – each one beat the last

glm 4.7 vs 5 vs 5.1 vs 5.2 – four models by @Zai_Org, same three tasks

context: @elonmusk and @jietang (founder of z ai, the lab behind glm) disagreed this week over how fast china's models are catching up to fable-class models

that argument got us thinking – so we decided to see for ourselves how much glm has improved recently. we took z ai's last four models:

• glm 4.7 – dec 22, 2025
• glm 5 – feb 12, 2026
• glm 5.1 – mar 27, 2026
• glm 5.2 – jun 13, 2026

and ran them all on @aimlapi on 3 simple prompts – animate water sloshing in a pool, a row of dominoes toppling, and a spaceship dodging meteors through space. all html, no libraries

we saw every version beat the one before it on every test. the output quality is simply better

and the reasoning time tells the story:

- pool sim: 31s on 4.7 → 362s on 5.2
- domino chain: 32s → 606s
- spaceship: 61s → 492s

it's not stalling – it's planning, checking itself, catching its own mistakes before it commits. and the output follows: 4.7 gave you a scroll animation, 5.2 lets you click the water and make waves

six months, four models, same three prompts. the gap between each one is visible to the naked eye – no benchmark needed. if glm is improving this fast, the gap probably closes sooner than any of us expect