GLM 5.1 Tops SWE-Bench Pro at 58.4% — First Chinese Model to Beat Claude and GPT on Real-World Coding

TamilTech AI summary

Z.ai’s GLM 5.1 just hit 58.4% on SWE-Bench Pro, edging past Claude Opus 4.6 and GPT-4.5 and becoming the first Chinese and first open-source model to lead that tough real-world coding benchmark. SWE-Bench Pro matters because it asks models to fix actual GitHub issues across big codebases, not toy problems, so a nearly six-out-of-ten success rate is a serious signal for autonomous software work. The model is a 744B MoE release under an MIT license, trained on Huawei Ascend chips rather than NVIDIA, and Z.ai also claims it can run multi-hour coding sessions on its own—though that longer autonomy piece is still mostly self-reported. For developers it offers near-frontier coding quality at a much lower price than Claude or GPT, plus the freedom to self-host or fine-tune without vendor lock-in. Keep the caveats in mind: independent checks are still catching up on some claims, the model is huge to run yourself, and you should test it on your own stacks before you fully lean on it.

• What is GLM 5.1?
• Does GLM 5.1 really beat Claude Opus?
• Can I use GLM 5.1 for free?
• What is special about GLM 5.1 being trained on Huawei chips?

AI-assisted summary, checked by the TamilTech editorial team.