GLM-5

Here is the source from Z.ai themselves! https://z.ai/blog/glm-5?_gl=1*qx9wgd*_gcl_au*MTAwMTgwMTkxMy4xNzcwODMwMjY0*_ga*MTMzOTcyODAxOS4xNzcwODMwMjY0*_ga_Z8QTHYBHP3*czE3NzA4MzAyNjMkbzEkZzEkdDE3NzA4MzAzMjIkajEkbDAkaDA.#:~:text=35.4-,Humanity%27s%20Last%20Exam,w/%20Tools,-50.4 It just barely made the top 5 HLE scores though. Surely Deepseek V4 will set the record soon. -Claude Opus 4.6 best score as reported by them is 53.1 as current leader. -Next up is SupAI with an ensemble structure at 52.15. -Then we have Moonshot - Kimi K2-Thinking-0905 coming in at 51% -Even Grok 4 got 51% -So, yes it slide in to the top 5 at 50.4%, which shows we are making progress and it might show promise.

GLM-5 is the latest flagship from Z.ai, and after putting it through a layered stress test on Yodayo, I wanted to give you my thoughts (like I do from time to time lol). IMPORTANT DISCLAIMER: I ran this on Yodayo (aka Moescape) as a hosted model there. From what I can tell, the <think> function is suppressed, as it is significantly faster output compared to running via proxy elsewhere. Also, Yodayo has its own internal economy with a credit system that you can earn buy, sub, get daily, etc. You spend for messages, but direct cost is hard to compare because you get message and reply at a flat rate on Yodayo. Elsewhere, paying per token, you pay for input and output both, and context can ramp up as your chat increases PLUS a model like GLM 5 outputting consistent Chain of Thought will chew tokens. TLDR, for a pure cost analysis, this is only going to mention Yodayo's internal ecosystem. But in general, proxy will likely be cheaper on paper, but mileage can vary greatly with practical use and user habits. As for the mode's capability, the rest of this can apply to any instance, and with thinking going it is usually very good at prompt adherence. Anyway, back to the show! The scenario I used wasn’t light work. It involved shared history, bruised egos, mentor-student tension, social pressure in a classroom setting, combat choreography, and a pivot into dominance-charged intimacy. Not just “fight scene go brrr” but emotional continuity across escalating turns. Basically, there's a shared past, a challenge, layered feelings underneath, and I chose to escalate the opening conflict set up by the creator into a fight for dominance. GLM-5 handled it with confidence. That’s the word that kept coming to my mind. Confidence. It committed to narrative choices. It escalated when escalation made sense. It didn’t stall or hedge. It kept up the rival tension without too much melodrama. Even under a fairly constraint-heavy prompt (my own tweaked prompt and settings), it stayed disciplined - no user agency bleed, no random fourth wall cracks, no rushed resolution. To give it context, I ran the same chat with Claude Haiku and GLM-4.7. Haiku, as expected, played it well. It delivered similar quality and story flow with slightly less melodrama, but more restraint. If anything, Haiku felt safer, composed. But it also took fewer risks. GLM-5 showed more narrative aggression and character agency. Haiku was the veteran assassin. GLM-5 was the heavyweight striker. Ultimately, I see GLM-5 and Haiku as side-grades, quality-wise. YMMV depending on what you prefer. Haiku is the more cost-efficient choice. GLM-5 brings sharper scene propulsion. I'd argue, GLM-5 feels more alive. GLM-4.7 was the most interesting comparison. Under the same exact prompt, 4.7 carried the chat in a noticeably different direction. Its webnovel-style dramatic bias came out much stronger at the start - {{char}} was a lot more volatile out the gate. Then, when things got more intense, it pivoted hard into power dynamics and effectively deflated the rivalry tension instead of sustaining it. Where GLM-5 and Haiku dove into the fight, 4.7 ultimately backed away from it. {{char}} backed down and turned the boil into a simmer, moving the chat more into a bit of a cold war between us. It is a little hard to explain, and some of this may be my own prompt, which does urge for nuanced depictions that dont over-rely on genre tropes. But my point is that GLM 5 picked a lane and committed, but rolled with with the direction of the scene and let {{char}} evolve. 4.7 seemed to start off wanting to sprint, and then caught itself and decided to pace in the opposite direction. 4.7 is a model I have high praise for, so this isn't a dig - just noting I think 5 brings with it a lot more decisive scene choices. Overall, I think GLM-5 is a strong offering on Yodayo. Whether it’s worth the premium tier is up to the user, but it absolutely competes in that space. It feels more assertive than 4.7 and capable of subbing in for higher tier models in confrontation-heavy RP without feeling outclassed. It may be a hard sell to a Haiku fan, but it is worth considering for a Sonnet user, if that makes sense. One last note - if Yodayo ever enables visible Chain of Thought reasoning for GLM-5 the way they do for DeepSeek R1 or Gemini Pro, I think it would elevate even further. In off-platform testing, GLM-5 is exceptionally good at following a well-tuned prompt and adapting tightly to user preference. It rewards careful engineering.

Because reliability and throughput are part of the daily discussion in this sub, I took it upon myself to create a small dashboard app to improve transparency. The dashboard tracks basic metrics such as tokens per second and time to first token for GLM-5 and GLM-4.7 once every hour. I additionally compare them to the performance of the standard API (let’s see for how long... API costs add up, after all). When discussing performance, it’s important to mention that I'm on the Pro plan. All metrics are, of course, only approximations and certainly not a dependable benchmark. Anyways try it here if you are interested: https://zaimonitor.vercel.app/