Zhipu released GLM-5.3 on August 14, and the API went live around August 18: $1.40 per million input tokens, $4.40 per million output tokens, with a 1-million-token context window. Independent evaluator Artificial Analysis gave the model an Intelligence Index score of 60 that same day, ranking it 8th among the 182 models it tracks. The median score for reasoning models in the same price tier is 35.
The real story here has two layers: the score, then the price. Lining up several open-weight or publicly accessible models from the same round of Artificial Analysis data tells a fuller picture. GLM-5.3 scores 59.5 at its top configuration (rounded to 60 on the model page), costs $0.68 per task, and outputs 41,107 tokens per task. Kimi K3 scores 59.7, costs $0.84, and outputs 25,474 tokens. Qwen 3.8 Max scores 58.1, costs $1.13, and outputs 38,287 tokens. xAI's Grok 4.6 (high) scores 60.9, costs $0.84, and outputs 21,735 tokens. GLM-5.3 trails Kimi K3 by just 0.2 points, putting the two roughly tied at the top of the open-weight pack — but GLM-5.3 needs about 60% more tokens to finish the same task, eating into its per-token price advantage. Even so, its per-task cost of $0.68 still comes out lowest in the group.
On launch day, Zhipu emphasized coding and security
In its August 14 release notes, Zhipu described GLM-5.3 as the product of "extreme post-training" on the same base model as GLM-5.2 — the underlying base didn't change, but training time and environment were substantially extended. The benchmark results it highlighted cluster around coding and agent tasks: Terminal Bench 3.0 rose from GLM-5.2's 4.6 to 28.3, DeepSWE v1.1 rose from 46.2 to 66.9, and Agents' Last Exam rose from 23.8 to 28.5. On Zhipu's own Z.ai Code Bench (High), the model hit 31.4% accuracy, ahead of the 29.5% it lists for Claude Opus. Zhipu's internal evaluation puts the coding improvement over GLM-5.2 at 50%.
"GLM-5.3 is the open-source model with the strongest programming ability."
That's Zhipu's own positioning: the strongest coding model in the open-source camp.
The other capability singled out was cybersecurity. Zhipu said post-training produced "beyond expected" emergent security capabilities, with red-team testing turning up a cumulative 2,436 vulnerabilities. On CyberGym, the model scored 84.5%, close to Anthropic's Mythos 5 at 83.8%. On ExploitBench, it scored 54.4%, well behind Mythos 5's 78.0%. Zhipu says the coverage extends beyond software to internet protocols and robotic systems.
Weights land within two weeks, coding tools get first access
The open-source plan: weights go up on Hugging Face within two weeks of release (roughly by August 28), contingent on completing safety evaluation and hardening; license terms haven't been announced yet. Until then, the model is available through the API and a set of coding tools. Zhipu named ZCode and AutoClaw as its own integrations, alongside third-party coding platforms TraeWork, WorkBuddy, Qoder, and CatPaw.
Discussion on Hacker News around the Artificial Analysis data added a user-side perspective. Some developers said GLM-5.3's visible reasoning tokens are a plus — "you can stop it in time when it goes off track" — while others pointed out its token usage runs higher than Kimi K3's. For teams billed by the token, the $1.40/$4.40 sticker price needs to be multiplied by a longer output before it can be compared fairly; the price list alone is misleading.
What it means for developers in China
Placed in the broader pricing history of domestic models: when GLM-5.2 launched, Zhipu's pitch was that its coding scores approached closed-source models at a sixth of GPT-5.5's price. GLM-5.3 pushes the absolute score up another notch while keeping pricing in the same range. For developers building in China, the real comparison points are the API prices of DeepSeek V4 Pro and Kimi K3, plus each vendor's peak/off-peak discounts. GLM-5.3 has now entered that scoring conversation — whether a price war follows depends on what happens once the weights are released in two weeks and third-party inference platforms start setting their own rates.
Sources: Zhipu GLM-5.3 release notes, Artificial Analysis model page, CocoLoop, Hacker News discussion, VentureBeat; verified API input/output pricing, Intelligence Index scores and rankings, per-task cost and output-token figures (Artificial Analysis's unified evaluation environment), official benchmark numbers, and the weight-release timeline.