An open-weights release from Z.AI has landed a competitive mark on a benchmark designed to probe genuine ML engineering intuition rather than rote patterns, slotting between two leading closed models while the updated test suite now factors in per-run API costs across nearly two dozen tasks.
Yup, this confirms it.
Open-models have now surpassed Opus 4.5
GLM 5.2 (high) scores 67.3% on WeirdML, a score between Opus 4.5 and Gemini 3 Pro.
This is a much higher score than I expected, and GLM 5.2 max (still running) could score even better.
It looks like a very solid model!
GLM 5.2’s result surprised the benchmark creator, who called the score higher than expected and the model very solid, underscoring that publicly available weights can now match or approach frontier performance on tasks requiring iterative code debugging under fixed compute limits.
With v2 tracking costs and expanding to 19 tasks, the leaderboard shows no single model dominating every point on the accuracy-versus-price curve, leaving open how developers will weigh GLM 5.2’s placement against alternatives once full cost data and the still-running higher variant become available.
Many users are excited about GLM 5.2 reaching or surpassing Opus 4.5 performance on WeirdML because rapid open-source gains signal the frontier is closing fast.
No Digg Deeper questions have been answered for this story yet.
China breaks into the frontier
GLM 5.2 (high) scores 67.3% on WeirdML, a score between Opus 4.5 and Gemini 3 Pro.
This is a much higher score than I expected, and GLM 5.2 max (still running) could score even better.
It looks like a very solid model!
> (high) 67.3% Assuming their scaling on other SWE evals, (max) still lands somewhere around my lower bound, but I might not have to eat an L today. Will Havard be shocked? Also remains to be seen! GLM-5 series seem to show the fastest improvement for one model's post-trains ever
GLM 5.2 (high) scores 67.3% on WeirdML, a score between Opus 4.5 and Gemini 3 Pro.
This is a much higher score than I expected, and GLM 5.2 max (still running) could score even better.
It looks like a very solid model!
This is just high… GLM recommends max for basically everything. Holy moly.
GLM 5.2 (high) scores 67.3% on WeirdML, a score between Opus 4.5 and Gemini 3 Pro.
This is a much higher score than I expected, and GLM 5.2 max (still running) could score even better.
It looks like a very solid model!
8.9% between GLM-5 and GLM-5.1. 10.2% between 5.1 and 5.2. Remarkably, lower cost (a pittance really); same LOC count. Havard's prior was reasonable, the difficulty increases. But now the absolute frontier is within sight. 1 more such jump to opus-4.7 (high). 1 more to Fable.
> (high) 67.3% Assuming their scaling on other SWE evals, (max) still lands somewhere around my lower bound, but I might not have to eat an L today. Will Havard be shocked? Also remains to be seen! GLM-5 series seem to show the fastest improvement for one model's post-trains ever
@htihle thanks for the points but the shock doesn't count in my book until we see ≥0.68 On a related note, I think they'll work next on improving CoT scaling. Their current curves are sad, like those of Opus; they must want OpenAI's.
@scaling01 So what will happen to all the ai labs who are charging crazy prices to keep their company afloat when open source eventually wins, GLM is really good
I think 5.4(xhigh) is a very trashy result. bloated code, bloated reasoning, bloated price. If GLM holds this pace and delivers 5.3 to "merely", say, 0.74 without bloat – I'll deem this a better result. GPT 5.2 (xhigh) was the real start of RSI. That'll do for the next spin.
8.9% between GLM-5 and GLM-5.1. 10.2% between 5.1 and 5.2. Remarkably, lower cost (a pittance really); same LOC count. Havard's prior was reasonable, the difficulty increases. But now the absolute frontier is within sight. 1 more such jump to opus-4.7 (high). 1 more to Fable.
@teortaxesTex I am pretty shocked! Bayes points to you.
@scaling01 this Fable delay is going to open the door for open source
@scaling01 It's still crazy to think that we're getting Opus 4.5 level open-source model
Note that it's on the Pareto frontier, by a good margin. Not a single model cheaper scores anything like that. They have a safety margin. I predict they will improve CoT scaling curve in 5.3, maybe add Pro/xhigh, to get something like extra 10% for 5x more tokens.
I think 5.4(xhigh) is a very trashy result. bloated code, bloated reasoning, bloated price. If GLM holds this pace and delivers 5.3 to "merely", say, 0.74 without bloat – I'll deem this a better result. GPT 5.2 (xhigh) was the real start of RSI. That'll do for the next spin.
@zephyr_z9 My market analysis is as follows:
👇Strategy as follows
👇
👇
👇
@scaling01 the closed model lead is shrinking faster than most expected
open models aren't catching up because of one thing but because of compounding gains across multiple vectors this shifts the economics for how companies build and ship models
@zephyr_z9 My market analysis is as follows:
👇Strategy as follows
👇
👇
👇
@poffyit @scaling01 How is open source going to win?
@zephyr_z9 My market analysis is as follows:
👇Strategy as follows
👇
👇
👇
@zephyr_z9
@zephyr_z9 The only way to play this trade is to long nebius in my opinion.
@poffyit @scaling01 GLM = goblin language model in 2026😂
@teortaxesTex 这进化速度确实有点超预期了