Benchmark Performance

Artificial Analysis Intelligence Index v4.1

GLM-5.2 scores 51 on the Artificial Analysis Intelligence Index v4.1 — the highest score ever recorded by an open-weight model . This places it 4th overall on the full leaderboard, behind only Claude Fable 5 (60), Claude Opus 4.8 (56), and GPT-5.5 (estimated ~53-55 range) . It beats other Chinese open-weight rivals: MiniMax-M3 (44), DeepSeek V4 Pro Max (44), and Kimi K2.6 (43) . The version-over-version delta from GLM-5.1 (40) to GLM-5.2 (51) is +11 points — a larger jump than most minor-version releases .

On Artificial Analysis's Coding Index, it scores 68.8, outperforming 99% of tracked models . The model also sits on the Pareto frontier of the Intelligence vs. Cost-per-Task chart, meaning no other open-weight model offers this capability at a lower cost per task .

Key Coding and Reasoning Benchmarks

According to independent evaluation company Vals AI, GLM-5.2 performed best among all models (including closed-source) on five benchmarks: Vals Index, Harvey's Legal Agent Benchmark, Finance Agent v2, ProofBench, and Vibe Code Bench .

Architecture

GLM-5.2 retains the same 744B total / 40B active Mixture-of-Experts (MoE) design as its predecessor GLM-5.1, but introduces two major architectural innovations :

• IndexShare — A novel sparse-attention optimization that reuses a single lightweight indexer across every four Sparse Attention layers, reducing per-token FLOPs at 1M context by approximately 2.9× .
• Improved Multi-Token Prediction (MTP) — Enhanced speculative decoding layer increases acceptance length by up to 20%, reducing latency on long outputs .

Key Specs

• Context window: 1 million tokens (usable, not just theoretical)
• Maximum output: 131,072 tokens per generation
• Thinking modes: "High" and "Max" effort levels for deeper reasoning
• License: MIT (fully open-weight)

Cost

API Pricing (per 1M tokens, from Z.AI's official docs)

Competitive Context

• GLM-5.2 costs roughly 1/6 the per-token price of GPT-5.5 and Claude Opus 4.8 for comparable or better coding performance .
• Also available via subscription: GLM Coding Plan tiers at ~$10/month (Lite), ~$30/month (Pro), and ~$80/month (Max), billed quarterly .
• Z.AI is on the Pareto frontier for Intelligence vs. Cost per Task .

Tech CEOs and Industry Praise

• Guillermo Rauch (CEO of Vercel) — Posted he was "genuinely impressed, almost shocked" at GLM-5.2's coding ability, adding the model "changes things" .
• Elon Musk — Liked social media posts about the model, signaling approval .
• Aaron Levie (CEO of Box) — Expressed strong positive reaction to the model's capabilities, describing what's happening with open weights AI as "pretty remarkable" .
• Hugging Face, the largest open-source AI community, personally funded six hours of free global compute for GLM-5.2 inference, a first for the platform .
• The Western tech community broadly described the week of its release as "one of the best weeks of all time" for open-source AI .

Strategic and Geopolitical Context

U.S. Entity List and Export Controls

Z.AI (Zhipu AI) was added to the U.S. Entity List in January 2025, restricting its access to American advanced semiconductors and chip-making equipment . The release of GLM-5.2 came just after the U.S. government ordered the shutdown of Anthropic's Claude Fable service, a direct competitor . Z.AI founder Jie Tang explicitly referenced that shutdown in his announcement, stating: "We deeply regret the sudden restrictions on certain frontier models. The sudden denial of access to frontier models for non-technical reasons strengthens our conviction that 'science should be global'" . On June 15, 2026, Zhipu's stock surged 33% as Wall Street raised bets on Chinese AI companies after Washington tightened restrictions on foreign access to advanced U.S. models .

Hong Kong IPO

Zhipu AI completed its Hong Kong Stock Exchange IPO on January 8, 2026, raising approximately HKD 4.35 billion (~$558 million) at a market capitalization near $52.83 billion . The company is led by CEO Zhang Peng.

Chinese Open-Weight Models vs. Western Frontier AI

GLM-5.2 is the clearest signal yet that Chinese open-weight models have closed the gap with Western proprietary frontier models. On several long-horizon coding benchmarks it now beats GPT-5.5 outright and sits within 1% of Claude Opus 4.8 . The model demonstrates that export controls have not slowed Chinese AI labs' ability to innovate on architecture (IndexShare) and achieve competitive benchmark scores . Z.AI's strategy combines open-weight release (MIT license), aggressive pricing (~1/6 the cost of frontier closed models), and a focus on long-horizon agentic coding tasks — directly targeting the developer audience that Western labs have dominated . The release triggers a broader competitive dynamic: if Chinese open-weight models stay on this trajectory, the differentiation Western labs charge for closed-source access becomes harder to justify, especially on cost-sensitive coding and agent workloads .

Benchmark Performance

Artificial Analysis Intelligence Index v4.1

GLM-5.2 scores 51 on the Artificial Analysis Intelligence Index v4.1 — the highest score ever recorded by an open-weight model . This places it 4th overall on the full leaderboard, behind only Claude Fable 5 (60), Claude Opus 4.8 (56), and GPT-5.5 (estimated ~53-55 range) . It beats other Chinese open-weight rivals: MiniMax-M3 (44), DeepSeek V4 Pro Max (44), and Kimi K2.6 (43) . The version-over-version delta from GLM-5.1 (40) to GLM-5.2 (51) is +11 points — a larger jump than most minor-version releases .

On Artificial Analysis's Coding Index, it scores 68.8, outperforming 99% of tracked models . The model also sits on the Pareto frontier of the Intelligence vs. Cost-per-Task chart, meaning no other open-weight model offers this capability at a lower cost per task .

Key Coding and Reasoning Benchmarks

According to independent evaluation company Vals AI, GLM-5.2 performed best among all models (including closed-source) on five benchmarks: Vals Index, Harvey's Legal Agent Benchmark, Finance Agent v2, ProofBench, and Vibe Code Bench .

Architecture

GLM-5.2 retains the same 744B total / 40B active Mixture-of-Experts (MoE) design as its predecessor GLM-5.1, but introduces two major architectural innovations :

• IndexShare — A novel sparse-attention optimization that reuses a single lightweight indexer across every four Sparse Attention layers, reducing per-token FLOPs at 1M context by approximately 2.9× .
• Improved Multi-Token Prediction (MTP) — Enhanced speculative decoding layer increases acceptance length by up to 20%, reducing latency on long outputs .

Key Specs

• Context window: 1 million tokens (usable, not just theoretical)
• Maximum output: 131,072 tokens per generation
• Thinking modes: "High" and "Max" effort levels for deeper reasoning
• License: MIT (fully open-weight)

Cost

API Pricing (per 1M tokens, from Z.AI's official docs)

Competitive Context

• GLM-5.2 costs roughly 1/6 the per-token price of GPT-5.5 and Claude Opus 4.8 for comparable or better coding performance .
• Also available via subscription: GLM Coding Plan tiers at ~$10/month (Lite), ~$30/month (Pro), and ~$80/month (Max), billed quarterly .
• Z.AI is on the Pareto frontier for Intelligence vs. Cost per Task .

Tech CEOs and Industry Praise

• Guillermo Rauch (CEO of Vercel) — Posted he was "genuinely impressed, almost shocked" at GLM-5.2's coding ability, adding the model "changes things" .
• Elon Musk — Liked social media posts about the model, signaling approval .
• Aaron Levie (CEO of Box) — Expressed strong positive reaction to the model's capabilities, describing what's happening with open weights AI as "pretty remarkable" .
• Hugging Face, the largest open-source AI community, personally funded six hours of free global compute for GLM-5.2 inference, a first for the platform .
• The Western tech community broadly described the week of its release as "one of the best weeks of all time" for open-source AI .

Strategic and Geopolitical Context

U.S. Entity List and Export Controls

Z.AI (Zhipu AI) was added to the U.S. Entity List in January 2025, restricting its access to American advanced semiconductors and chip-making equipment . The release of GLM-5.2 came just after the U.S. government ordered the shutdown of Anthropic's Claude Fable service, a direct competitor . Z.AI founder Jie Tang explicitly referenced that shutdown in his announcement, stating: "We deeply regret the sudden restrictions on certain frontier models. The sudden denial of access to frontier models for non-technical reasons strengthens our conviction that 'science should be global'" . On June 15, 2026, Zhipu's stock surged 33% as Wall Street raised bets on Chinese AI companies after Washington tightened restrictions on foreign access to advanced U.S. models .

Hong Kong IPO

Zhipu AI completed its Hong Kong Stock Exchange IPO on January 8, 2026, raising approximately HKD 4.35 billion (~$558 million) at a market capitalization near $52.83 billion . The company is led by CEO Zhang Peng.

Chinese Open-Weight Models vs. Western Frontier AI

GLM-5.2 is the clearest signal yet that Chinese open-weight models have closed the gap with Western proprietary frontier models. On several long-horizon coding benchmarks it now beats GPT-5.5 outright and sits within 1% of Claude Opus 4.8 . The model demonstrates that export controls have not slowed Chinese AI labs' ability to innovate on architecture (IndexShare) and achieve competitive benchmark scores . Z.AI's strategy combines open-weight release (MIT license), aggressive pricing (~1/6 the cost of frontier closed models), and a focus on long-horizon agentic coding tasks — directly targeting the developer audience that Western labs have dominated . The release triggers a broader competitive dynamic: if Chinese open-weight models stay on this trajectory, the differentiation Western labs charge for closed-source access becomes harder to justify, especially on cost-sensitive coding and agent workloads .