GLM-5.3 claims coding gains, but developers question whether Z.ai is gaming benchmarks

Z.ai, a Chinese frontier model developer, unveiled GLM-5.3 on Friday as an incremental refinement of its GLM-5.2 codebase, with performance gains achieved entirely through post-training optimization. The company positions the model as significantly more capable at tackling complex coding tasks and extended workflows.

Post-training optimization as the engine of improvement#

The performance jump relied on no architectural changes, only enhancements during the post-training phase—which encompasses reasoning alignment, supervised fine-tuning, and reinforcement learning from human feedback. Z.ai described its approach as layering more diverse training environments, broader production workflows, and increased compute spending onto the existing foundation.

According to the company, its training scenarios now reflect how engineering and research teams actually work, including challenges that might take days for a human engineer to solve. One illustrative example involved giving the model access to compute clusters, storage, documentation, and experiment logs—then asking it to diagnose infrastructure bottlenecks, implement fixes, run tests, and deliver measurable end-to-end speedups without breaking functionality.

Claims of dramatic coding improvement#

Z.ai introduced its own benchmark, Z.ai Code Bench, showing a 50% improvement over GLM-5.2—nearly double the prior version's performance. The company framed this homegrown metric as more representative of real-world use since it avoids contamination from public test sets. The results also covered several public benchmarks, including TerminalBench 3.0, DeepSWE, and OpenAI's GDPVal-AA v2.

Skepticism from real-world practitioners#

However, developer practitioners offering independent perspectives pushed back on these claims. Nishant Soni, co-founder of long-horizon AI engineering firm NonBioS.ai, expressed caution: "To my knowledge, no benchmarks objectively demonstrate the claimed superior long-horizon capability of GLM-5.3 in real world tasks."

Soni's team has conducted internal testing and observed striking output similarity between GLM/Kimi models and Claude, whereas competitors like Gemini and Grok show more diversity—suggesting, he believes, industrial-scale distillation of Anthropic's models rather than fundamental capability breakthroughs. "I suspect this is largely an effort to deflect from the model's true source of frontier capability," Soni told The New Stack.