GLM-5.2 vs. Claude Opus 4.8: prova de codi 3D

G, L, M, 5.2, just landed on hugging face with full open weights under an MIT license, and the numbers it's posting are forcing a real conversation. This is Z.A.I.'s new flagship, the Chinese lab formally called G-Poo, and the whole thing was trained on Huawei silicon. No Nvidia chips were involved in any of it. That detail alone shifted some assumptions in the industry. The architecture is a mixture of experts set up. Total parameter count sits around 753 billion, but only about 40 billion stay active per token. That selective routing keeps the inference cost

manageable on z.ai side and on yours. They paired it with a real, one million token context window, not the kind that falls apart after 200,000 tokens, plus output up to roughly 131,000 tokens in one shot. two reasoning effort modes ship with it, labeled high and max, so you can dial between speed and depth depending on the job. On terminal bench 2.1, GLM 5.2 scored 81.0, a huge jump from GLM 5.1 at 62.0. On SWE bench pro it landed at 62.1, while GPT 5.5 sits at 58.6 on the same test and GLM 5.1 was at 58.4. On Frontier SWE, which measures long horizon coding work over many steps, GLM 5.2 hits 74.4 percent. GPT 5.5 is at 72.6.

Claude Opus 4.8, the model I'm comparing it against today, is at 75.1, so an open weight's Chinese model is sitting within one percentage point of Opus on that benchmark. And it costs roughly 1-6th as much per token through the official API. Index share is the architecture trick worth knowing about. It reuses one lightweight indexer across every four sparse attention layers, and the result is a 2.9 times reduction in per token compute at the full 1 million context. They also upgraded the multi-token prediction layer for speculative decoding, which bumps accepted token length by up to 20% translation, it stays fast when you push it into long sessions. CloudFableFive would be the natural comparison here, but FableFive got pulled. It launched on June 9 as Anthropics mythos class model, the tier above Opus, and got disabled

three days later. The US Commerce Department sent Anthropics an export control letter at 521pm Eastern on June 12. The directive barred any foreign national from accessing Fable 5 or Mythos 5, including foreign employees inside and Thropic itself. Anthropic couldn't enforce that filter in real time so they shut both models down for everyone. The trigger was a jailbreak demonstration the administration was worried about for cyber security reasons. Fables Out.

Opus 4.8 is what's left from the top tier. KimeK 2.7 code from Moonshot shipped right around the same time with a reported 21 point jump on their internal coding benchmark. Mini Max M3 is posting good numbers too. Quen 3.7 from Ali Baba keeps releasing variants. DeepSeek V4 is in the mix as well, but for the argument happening right now in the developer community, GLM 5.2 and Opus 4.8 are the two, everyone keeps comparing. So let me actually push them. I'm using both models through their official chat interfaces, GLM 5.2 on Z.Ayes Chat, Opus 4.8 on Claude.Aye,

same prompts on both sides, two builds. First test is a 3D Rubik's Cube using 3.js. Standard 3 by 3, colored stickers, click and drag to rotate any face, orbit camera with mouse, scramble button, reset button, and a move counter. Easy to describe,

harder to build cleanly in a single HTML file. Opus 4.8 returned a working cube. Stickers look clean, rotation feels stable. Scramble works, reset works, move counter ticks up properly. The aesthetic is classic and traditional. Flat colored stickers on a black plastic frame, plain dark background, no extras. It reads as a Rubik's cube, functional and tidy. G-L-M-5.2 took the same prompt and pushed it further. Rossi rounded stickers with a soft gradient on every face, a reflective sheen on top,

a thicker beveled frame, and a subtle ground shadow under the cube. It also added a keyboard control strip at the bottom labeling U, D, R, L, F, B. The six standard cube notation moves, plus a hint that holding shift gives counterclockwise turns. That's actual cubernalage baked into the interface. Opus didn't include any of that. Both cubes work. GLMs looks like someone who solved a real cube built the UI. Second test is a 3D endless runner called CubeRunner. 3.js, neon grid floor, player cube sliding left and right with arrow keys.

Incoming 3D obstacles, speed ramping up over time, score counter, particle effects on near misses, game over screen with restart. Single-H, TML file, no external assets. Opus 4.8 built a run-able game. A cyan cube on a purple grid, pink, and magenta obstacles, spawning ahead, a small, multiplier countertop right, showing speed. Controls respond, you can dodge, it plays. Visually though, it looks like a 3.js intro tutorial. Bear Grid, plain primitive shapes, no skybox, no background, no atmosphere. GLM 5.2 built something I'd actually screenshot for a portfolio.

A synth wave city scene with silhouetted skyscrapers, a giant purple sun setting on the horizon, a striped highway running into the distance, glowing pink and cyan neon walls on either side. The player is a faceted ice blue icosahedron, wrapped in a yellow ring. Obstacles vary, cones, pyramids, twisted knots. A flame effect sits in the middle of the lane As a hazard you have to swore around, the UI text at the bottom even reads drift between walls.

AD works too, meaning it added WASD fallback controls without being asked. Both games run. GLM's looks shipable. GLM 5.2 added details the prompt didn't request but a real developer would put in. Cube notation labels on the Rubik's build. WASD as a control fallback on the runner. backdrop instead of empty space. That kind of polish usually shows up on the second or third prompt with most models. Here it showed up on the first try, both times. Opus 4.8 deserves fairness too.

Both outputs ran without bugs. Both were stable. Opus tends to produce cleaner, more conservative code, and that's not nothing. For a production refactor, for a deep debugging session. For code I'm handing off to a team, I might still pick Opus because the surface area is smaller and easier to review. But for a one-shot demo for a fast prototype for a first impression piece, GLM 5.2 is putting more on the screen. Opus 4.8 through a cloud consumer plan runs $20 a month and a PI pricing is roughly $5 per million input tokens and $25 per million output.

GLM 5.2 on the official hosted API is around $1.40 in and $4.40 out per million tokens. The Z.A.I. subscription tier starts near $12.60 a month. And if you have the GPU capacity, you can pull the weights from hugging face under MIT and run it on your own hardware with zero ongoing cost. For solo devs, indie teams, and anyone in a country where US API billing is friction, that math is a real shift. You don't have to commit to opus. You can run a frontier coding model locally if your hardware is there or rent it for pennies if it isn't. The license is permissive enough for commercial use without restriction.

GLM went from 5 to 5.1 to 5.2 in a handful of months. Kimmy shipped K 2.7 code. max pushed M3. Quinn keeps cycling. The open weight side is no longer trailing behind closed frontier models. On Frontier SWE, GLM 5.2 sits within 1% of Opus 4.8 while costing about 82% less per token. That's not a gap I expected to see this year, and it's reshaping how I plan my workflow. Opus 4.8 still has its place in my stack, reasoning heavy debugging, long conversation and refactors, anything where I want the most conservative answer in the room. But for fast prototyping, front and work, anything visual or three-dimensional, GLM5.2 is now my default. The cube runner test alone sold me on that.

Alright, so, that's it from the video and I hope you enjoyed it if you did, please like this video and subscribe to the channel and I'll see you in the next video.