Z.ai's Glm 5.2 beats GPT-5.5 on coding tests
Z.ai has released glm 5.2, a 753-billion-parameter open-weights model built for long-horizon autonomous coding and engineering tasks. It is available now on Hugging Face, the Z.ai API, and more than 20 third-party coding environments.
The model also arrives with a 1-million-token context window and core weights under an unrestricted MIT open-source license. For teams that want to download it, customize it, fine-tune it, or run it locally or through virtual machines, that is a different distribution model from the usual cloud-only route.
GLM-5.2 and GPT-5.5
On SWE-bench Pro, GLM-5.2 scored 62.1, ahead of GPT-5.5 at 58.6 and GLM-5.1 at 58.4. That gap is small in absolute terms, but it puts Z.ai's newer model ahead of both the newer rival and its own predecessor on a coding benchmark that tests long-horizon work.
On FrontierSWE Dominance, GLM-5.2 scored 74.4%, compared with GPT-5.5 at 72.6% and Claude Opus 4.8 at 75.1%. On MCP-Atlas, it scored 77.0, above GPT-5.5 at 75.3 and just below Claude Opus 4.8 at 77.8.
IndexShare in GLM-5.2
Z.ai says IndexShare reuses the identical indexer across every four sparse attention layers, and says that at the maximum 1-million-token context length it reduces per-token compute FLOPs by 2.9 times. In practice, that points to a model that is trying to hold more context without paying the full compute cost every time the window stretches to its limit.
The model also includes an upgraded Multi-Token Prediction layer for speculative decoding, which Z.ai says boosts accepted token length by up to 20% during inference. It also offers selectable Thinking Modes called Max and High, which looks aimed at letting users trade speed, depth, and cost depending on the coding task.
Z.ai pricing and access
Enterprise subscription tiers for GLM-5.2 start at $12.60 per month. That low entry price is the friction point in this release, because the benchmark lead is only useful if the model's quality, licensing, and deployment options hold up in real codebases.
For enterprises, the practical question is whether open weights and local deployment outweigh the need to validate the model against proprietary systems in their own environments. Z.ai has put out a model that can be downloaded, modified, and run more broadly than a closed API service, and the benchmark numbers say it is competitive enough to warrant that test.
