1. A Big Jump in Coding Performance
Coding is the clearest area of improvement.
In Zhipu AI’s internal coding evaluations, GLM-5.3 scores about 50% higher than GLM-5.2.
The gap is particularly large on Terminal-Bench 3.0, where the score rises from 4.6 to 28.3, ahead of Kimi K3 at 17.4.
On DeepSWE, a benchmark focused on longer software-engineering tasks, the score moves from 46.2 to 66.9, putting it close to Claude Fable 5 at 69.7.
These aren’t just short code-completion tests. They involve working in terminals, understanding projects, editing across files, and keeping a task moving over multiple steps. That is where most of the improvement shows up.
2. Lower Token Usage
For coding agents, accuracy is only half the story. Token consumption can quickly become the more expensive part of a long task.
At the High reasoning setting, GLM-5.3 uses about 50,000 tokens per task on average while reaching 31.4% accuracy. Claude Opus 4.8 uses roughly 120,000 tokens for 29.5% accuracy in the same comparison.
That difference matters once an agent starts looping through files, running tools, revising code, and carrying context across many turns.
For heavy users, lower token usage translates directly into lower operating cost.
3. Stronger Cybersecurity Performance
Security is another area where the model posts substantial gains.
On the CyberGym vulnerability-reasoning benchmark, GLM-5.3 scores 84.5%, compared with 83.8% for Anthropic Mythos 5.
In a separate security effort, the model was used across 269 projects and helped identify 2,436 real vulnerabilities, including a high-severity DNS protocol issue that had remained undetected for more than 40 years.
That capability also raised the bar for releasing the weights safely. Zhipu AI delayed the open-weight release by two weeks to add further safeguards.
4. What Changed From GLM-5.2
The main change is in training, not model size.
Zhipu AI put more reinforcement-learning effort into terminal use, complex codebases, and long-horizon engineering tasks. Instead of training primarily around isolated prompts, the model spends more time working through environments that look closer to actual development work.
The gains are concentrated in the same areas: sustained execution, cross-file reasoning, and more involved engineering tasks.
GLM-5.3 is less a new foundation model and more a proof of how much post-training can still pull out of an existing one.
Comments (0)