Zhipu GLM-5.3 Coding Capability Surpasses Claude Opus 4.8 but Still Trails Fable 5
Nashnova编辑部
Zhipu (Z.ai) is set to release GLM-5.3, which scores 31.4% on a coding benchmark — beating Claude Opus 4.8's 29.5% — but Anthropic's flagship Fable 5 still leads at 39.5%, leaving a clear gap between China's open-weight contenders and the frontier.
31.4% vs 29.5% — where does GLM-5.3 win?
In high-effort mode, GLM-5.3 scores 31.4% on the coding benchmark using roughly 50,000 output tokens.
On the same benchmark, Anthropic's Claude Opus 4.8 scores 29.5% but consumes 120,000 tokens — more than double.
This means → GLM-5.3 not only scores higher but does so with far better token efficiency — fewer words, better work.
How far behind Fable 5?
Anthropic's flagship Claude Fable 5 reaches 39.5% in max-effort mode — 8.1 percentage points above GLM-5.3.
In plain terms = GLM-5.3 just cleared the Opus 4.8 bar, but the true ceiling is still a tier away.
This reflects the shape of the AI coding race: gaps within the top tier are narrowing, yet a structural gulf between flagships and challengers persists.
Same base, fast iteration — what is Zhipu's playbook?
GLM-5.3 is built on the same 700-billion-parameter foundation model as its predecessor, released in June this year — no new base training.
This means → Zhipu is betting on rapid fine-tuning atop a fixed base, trading lower cost for faster capability gains.
The approach is emblematic of China's open-weight camp chasing Anthropic and OpenAI. Whether GLM-5.3 can continue closing the gap with Fable 5 will be a key marker for the domestic large-model coding race.
Content is for reference only, not financial advice.