GLM-5.3 has moved quickly into the center of the open-model conversation. The AI model from Chinese startup Z.ai scores 60 points on the Artificial Analysis Intelligence Index, putting it level with Kimi K3 at the top among open models.
The headline is not only the ranking. GLM-5.3 also improves sharply over GLM-5.2, posts major gains on agentic tasks, and is priced below Kimi K3 on a per-task basis. At the same time, the release of its open weights is being held back while Z.ai tightens access around security-sensitive capabilities.
A new top score among open models
Artificial Analysis gives GLM-5.3 a score of 60 points on its Intelligence Index. That result ties GLM-5.3 with Kimi K3 for the top spot among open models.
The comparison with Z.ai's earlier model is also important. GLM-5.3 is seven points ahead of GLM-5.2, showing that the new release is not just a marginal update in the ranking used by Artificial Analysis.
For developers and companies tracking open AI models, that matters because model choice often depends on a mix of benchmark performance, availability, and cost. GLM-5.3 now appears in that discussion as both a leading open model by score and a direct successor with a clear reported improvement over GLM-5.2.
Agentic tasks drive the biggest improvement
The largest reported gains for GLM-5.3 come from agentic tasks. On the GDPval-AA v2 benchmark, its Elo score rises from 1,524 to 1,770.
That is a gain of 246 points. In the ranking cited by Artificial Analysis, it puts GLM-5.3 in second place on that benchmark, behind only Claude Opus 5, which is listed at 1,855.
Agentic tasks are especially relevant because they point to how a model handles multi-step work. The source does not describe the individual tasks inside GDPval-AA v2, so the safest reading is limited to the reported benchmark movement: GLM-5.3 makes its most visible jump in this area, and that jump is large enough to place it near the top of the cited comparison.
The improvement also changes how GLM-5.3 is positioned against its predecessor. GLM-5.2 remains cheaper per task, but GLM-5.3 brings stronger reported performance, especially where the benchmark emphasizes agentic behavior.
Cost puts pressure on rivals
Artificial Analysis pegs GLM-5.3 at $0.68 per task. That makes it more expensive than GLM-5.2, which is listed at $0.44 per task.
The increase is described as 1.5 times the cost of GLM-5.2. But GLM-5.3 is still priced below Kimi K3, which is listed at $0.84 per task.
On that comparison, GLM-5.3 is 19 percent cheaper than Kimi K3. That price gap matters because the two models are tied at 60 points on the Artificial Analysis Intelligence Index, according to the source.
The result is a sharper tradeoff for users evaluating these models:
- GLM-5.3 matches Kimi K3 on the cited open-model ranking.
- GLM-5.3 costs more per task than GLM-5.2.
- GLM-5.3 costs less per task than Kimi K3.
- GLM-5.3 shows its biggest reported improvement on agentic tasks.
That does not automatically make it the right model for every use case. The source gives benchmark and price data, not deployment results across real products. But it does show why GLM-5.3 is likely to draw attention from teams weighing open-model performance against operating cost.
Open weights are delayed
GLM-5.3 is accessible through Z.ai's API. However, the open weights are not being released immediately.
Z.ai is delaying the release of the open weights by about two weeks. According to the company, the reason is that GLM-5.3 is highly effective at detecting security vulnerabilities.
Before full access is made broader, Z.ai is strengthening controls and restricting full access to select security partners. The source does not give more detail about those controls or identify the partners, so the practical takeaway is limited but significant: API access is available, while open-weight access is delayed for security-related reasons.
That creates a split launch. Developers can use the model through the API, but those waiting for open weights will have to wait while Z.ai completes the additional access controls it says are needed.
What the ranking means now
GLM-5.3's position is notable because it combines three facts in one release: a 60-point Intelligence Index score, a 246-point Elo gain on GDPval-AA v2, and a per-task cost below Kimi K3.
Those facts make it a serious contender in the open-model field, at least according to the benchmarks and prices cited by Artificial Analysis. The delayed open-weight release keeps the story from being a straightforward availability win, but it also shows that Z.ai is treating the model's vulnerability-detection strength as a deployment issue, not only a benchmark result.
For now, GLM-5.3 stands as a high-scoring open model available through Z.ai's API, with open weights expected later after a delay of about two weeks. Its next test will be whether that combination of performance, cost, and controlled access translates into broader adoption once the open weights are released.