Teortaxes highlights how GLM-5.2 from Z.ai lifted its Artificial Analysis Intelligence Index score from 40 to 51, making it the strongest open-weights model on the composite while still trailing closed frontier entries like Claude Fable 5 at roughly 60 and Claude Opus 4.8 at 56. The same 744B-total model now ships with a 1M-token context window under an MIT license and shows its biggest lifts in scientific reasoning, agentic tasks, and coding benchmarks.
GLM 5.2 improved by 11 points over GLM 5.1 (and 5, somehow) The same magnitude of improvement would put V4 on par with Opus 4.8 or 5.5-xhigh – absolute frontier… that you can touch If we still don't have Fable/5.6 in two weeks, it'll be funny «Open source fully caught up!»
If Deepseek is only in PREVIEW Mode. Does the Official Release Move The Needle ?
Sub-metric gains include double-digit jumps on CritiPt, Humanity’s Last Exam, Terminal-Bench, and tau3 banking, plus higher token usage that signals more internal reasoning steps. These changes keep the model on the cost-performance frontier without any parameter increase.
The 11-point move narrows the visible gap on this index, yet top closed scores still sit 5–9 points ahead and many of those models remain unreleased or preview-only. Whether the same leap pattern appears in upcoming open releases like DeepSeek V4 stays an open variable.
Positive users highlight GLM 5.2's fast gains nearing frontier rankings while negative users criticize coding flaws in related models and question closed-model evaluations.
No Digg Deeper questions have been answered for this story yet.
@teortaxesTex Sorry, the goalpost is now “yeah so this unavailable closed model is better”
GLM 5.2 improved by 11 points over GLM 5.1 (and 5, somehow) The same magnitude of improvement would put V4 on par with Opus 4.8 or 5.5-xhigh – absolute frontier… that you can touch If we still don't have Fable/5.6 in two weeks, it'll be funny «Open source fully caught up!»
Damn, right, how could I forget again
@teortaxesTex Sorry, the goalpost is now “yeah so this unavailable closed model is better”
V4-Flash equaling GLM-5.2's score would also be neat (though I think it's far closer to its ceiling vs Pro) – "almost frontier" on consumer hardware. One cheap area of improvement for both is atrocious hallucination discipline sounds a bit too good to be true
GLM 5.2 improved by 11 points over GLM 5.1 (and 5, somehow) The same magnitude of improvement would put V4 on par with Opus 4.8 or 5.5-xhigh – absolute frontier… that you can touch If we still don't have Fable/5.6 in two weeks, it'll be funny «Open source fully caught up!»
don't laugh at Google! They do have an insane multimodality lead!
@aquiffoo pretty good
@teortaxesTex i believe V4 Pro max/high will score 53/50, V4 Flash will score 47/44
other estimates: - Qwen3.8 Max: 52 - Qwen3.8 Plus: 45 - Grok 4.5 high: 52 - GPT-5.6 Sol max: 62 - GPT-5.6 Terra max: 57 - GPT-5.6 Luna max: 50 - Gemini 3.5 Pro high: 56 - Claude Sonnet 5 max: 56
@teortaxesTex v4 pro is so dogshit at coding, surprised how far behind they're considering the model size against the other chinese labs
@teortaxesTex 这迭代速度有点吓人，开源模型真的快追平了
@teortaxesTex i dont think deepseek improves the way glm does.
I think GLM does something different. latest kimi didnt improve the way glm did either.
@teortaxesTex so we hyped?
@xeophon @teortaxesTex “That’s cool, can a 3rd party evaluate your fancy giga-model?”
“no”
@teortaxesTex 记忆总是断片真的好真实