Chinese AI company Z.ai launches model that challenges Anthropic and OpenAI

Chinese AI company Z.ai is introducing a new language model, GLM-5.3, that directly competes with the latest models from Anthropic and OpenAI. On an internal coding test by Z.ai itself, the model achieves a score of 31.4%, which is higher than the 29.5% of Anthropic’s Claude Opus 4.8. Claude Fable 5 still tops the list for now with 39.5%. Meanwhile, Z.ai’s share price has risen by 1.000% since its stock market debut in January.

In brief:

• Z.ai’s newest model GLM-5.3 outperforms Claude Opus 4.8 on coding tasks while also using fewer tokens.
• On 3 of the 9 benchmarks, GLM-5.3 scores higher than both Anthropic and OpenAI, namely on AutomationBench, GDPVal-AA v2 and CyberGym.
• Z.ai’s share price has risen by 1.000% since its IPO in January, although the price is still about 57% below the record level reached in June.

GLM-5.3 wins on efficiency and specific tasks

GLM-5.3 stands out because it beats its competitor Claude Opus 4.8 while using fewer output tokens per task. This makes the model not only more accurate in certain areas, but also more economical to use. On the internal Z.ai Code Bench, assessed via Claude Code 2.1.207, GLM-5.3 scores 31.4% at high effort, compared with 29.5% for Claude Opus 4.8.

Yet Claude Fable 5 remains the strongest in coding performance for now, with a maximum score of 39.5%. The older GLM-5.2, with a maximum score of around 23.4%, lags considerably behind both newer models.

Strong results across multiple benchmarks

Across nine benchmarks, including Terminal Bench 3.0, DeepSWE, AutomationBench and CyberGym, GLM-5.3 impresses in specific areas. On AutomationBench, the model achieves a score of 48.2, while competitors such as Kimi K3 (46.7), Mythos/Fable 5 (46.2) and GPT-5.6 Sol (45.8) stay below that. On GDPVal-AA v2, GLM-5.3 even scores 1,769 points, clearly more than the scores of the other models. On CyberGym it scores 84.5, compared with scores between 77 and 84 for the rest.

On other benchmarks such as DeepSWE (66.9) and HLE w/ Tools (62.5), GLM-5.3 also performs well, although the scores of GPT-5.6 Sol are slightly higher there, at 72.7 and 64.5 respectively.

Z.ai share surges after IPO

The launch of GLM-5.3 comes at a notable moment for Z.ai as a listed company. Since the IPO in January, the share price has risen by no less than 1.000%. At the same time, the price is still about 57% below the record level reached in June, which indicates that the share has fallen sharply after a strong rally.

With GLM-5.3, Z.ai is trying to establish itself as a serious competitor in the international AI race, at a time when companies such as Anthropic and OpenAI are calling the shots in the development of large language models.