GLM-5.2: The New Champion of Open Weights Models

Published June 17, 2026 | Source: Artificial Analysis

Z ai has officially released GLM-5.2, which has ascended to the top spot for open weights models on the Artificial Analysis Intelligence Index. With a score of 51, the model now defines the Pareto frontier when balancing Intelligence against Cost per Task.

While GLM-5.2 maintains the same architectural footprint as its predecessor—GLM-5.1 (featuring $744\text{B}$ total parameters and $40\text{B}$ active parameters)—it achieves a massive $\Delta 11$ point increase on the Intelligence Index v4.1. This leap places it significantly ahead of competitors like MiniMax-M3 (44) and DeepSeek V4 Pro (max) (44).

---

🚀 Key Performance Milestones

Summary: GLM-5.2 is currently the premier open weights model on the Intelligence Index v4.1, outperforming MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6.

Scientific & Reasoning Breakthroughs

The model shows substantial gains across nearly all evaluations, with the most dramatic improvements seen in scientific reasoning:

• CritPt: $\uparrow 16\%$ (reaching $21\%$)
• HLE: $\uparrow 12\%$ (reaching $40\%$)
• AA-LCR: $\uparrow 9\%$ (reaching $71\%$)
• tau3 banking: $\uparrow 15\%$ (reaching $27\%$)
• SciCode: $\uparrow 7\%$ (reaching $50\%$)
• TerminalBench v2.1: $\uparrow 16$ points (reaching $78\%$)
• GPQA Diamond: $\uparrow 3$ points (reaching $89\%$)

Agentic Capabilities: GDPval-AA v2

GLM-5.2 dominates the GDPval-AA v2 benchmark, which measures real-world agentic performance.

Model	GDPval-AA v2 Score
GLM-5.2	1524
MiniMax-M3	1418
DeepSeek V4 Pro (max)	1328
GPT-5.5 (xhigh reasoning)	1514

Note on GDPval-AA v2 Methodology: This updated benchmark improves upon the original by:

1. Baselining Elo to human performance at $1000$.
2. Implementing a rotating panel of frontier-model judges.
3. Extending the turn limit from 100 $\rightarrow$ 250 to better evaluate long-horizon agent trajectories.

---

📉 Efficiency and Cost Analysis

The Pareto Frontier

GLM-5.2 occupies a highly strategic position on the Intelligence vs. Cost per Task chart, offering the lowest cost for its specific intelligence tier.

Cost Comparison per Task:

• GLM-5.2: ~$0.46
• GLM-5.1: $0.25
• Kimi K2.6: $0.31
• MiniMax-M3: $0.18
• DeepSeek V4 Pro (max): $0.05

Token Consumption

One trade-off for this intelligence is token efficiency. GLM-5.2 is more "verbose" in its reasoning process:

• Total output tokens per task: $43\text{k}$ (of which $37\text{k}$ is dedicated to reasoning).
• Comparison: This is significantly higher than GLM-5.1 ($26\text{k}$), MiniMax-M3 ($24\text{k}$), Kimi K2.6 ($35\text{k}$), and DeepSeek V4 Pro ($37\text{k}$).

---

🛠️ Technical Specifications & Availability

Model Profile

• License: MIT
• Architecture: 744B Total / 40B Active Parameters
• Context Window: 200K $\rightarrow$ 1M tokens
• Omniscience Index: 4 (Up from 2 in GLM-5.1)
  • Accuracy: $25.1\%$ (vs $24.2\%$)
  • Hallucination Rate: $28.1\%$ (vs $29.4\%$)
  • Attempt Rate: $47\%$ (Flat)

Pricing Structure

The pricing remains consistent with the previous version:

{
  "pricing_per_1M_tokens": {
    "input": "$1.40",
    "output": "$4.40",
    "cache_hit": "$0.26"
  }
}

Where to Access

Beyond Z ai's own API, GLM-5.2 is available via:

• DeepInfra
• Novita
• Nebius
• Parasail
• Siliconflow
• GMI Cloud
• Baseten
• Fireworks

---

🖼️ Visual Data & Analysis

Figure 1: Intelligence Index Positioning

Figure 2: Cost vs. Intelligence Pareto Frontier

Figure 3: Detailed Evaluation Breakdown

Figure 4: Token Efficiency Analysis

Figure 5: Agentic Performance Metrics

Figure 6: Comparison with Proprietary Models

Figure 7: Intelligence Index v4.1 Distribution

Figure 8: Reasoning Token Breakdown

---

Further Reading:

• Explore GLM-5.2 at artificialanalysis.ai/models/glm-5-2
• Read about the shift toward agentic workloads in the Intelligence Index v4.1 announcement (June 16, 2026).
• Check out the launch of Claude Fable 5, the first Mythos-class model (June 10, 2026).