概要
GLM-5.2, released by Z.ai (formerly Zhipu AI) on June 16, 2026, is a 753B-parameter mixture-of-experts (40B active) foundation model purpose-built for long-horizon agentic coding and engineering tasks. It represents a generational leap over its predecessor GLM-5.1, expanding the context window from 200K to a solid 1M tokens while dramatically improving coding capabilities across every major benchmark. On Terminal-Bench 2.1 it scored 81.0 versus 63.5 for GLM-5.1, and on FrontierSWE it jumped from 30.5 to 74.4—within 1% of Claude Opus 4.8.
The model's positioning is deliberate: it targets the gap between expensive closed frontier models and less capable open alternatives. At $1.40/$4.40 per million tokens, it delivers near-frontier performance at roughly one-sixth the output cost of GPT-5.5 and Claude Opus 4.8. The MIT license enables full self-hosting and commercial use without regional restrictions. Key architectural innovations include IndexShare (reducing per-token FLOPs by 2.9× at 1M context) and improved multi-token prediction layers achieving 20% longer acceptance lengths for speculative decoding.
Community reception has been strong. Lambda Labs called it a "DeepSeek moment for agents," noting that experienced labs began replacing closed-source workloads with GLM-5.2 within weeks of release. DeepLearning.ai highlighted it as the top open-weights model for post-training benchmarks and web-development coding. The model's agentic RL training with anti-hack mechanisms—a system to detect and block reward-hacking behaviors during training—sets it apart in its ability to reliably complete extended autonomous tasks without shortcutting.
ベンチマーク＆性能
## Comprehensive Benchmark Results
### Reasoning & Knowledge
| Benchmark | GLM-5.2 | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| GPQA-Diamond | 91.2% | 93.6% | 93.6% | 94.3% |
| HLE (with Tools) | 54.7% | 57.9% | 52.2% | 51.4% |
| HLE (text only) | 40.5% | 49.8% | 41.4% | 45.0% |
| AIME 2026 | 99.2% | 95.7% | 98.3% | 98.2% |
| HMMT Feb 2026 | 92.5% | 96.7% | 96.7% | 87.3% |
| IMOAnswerBench | 91.0% | 83.5% | — | 81.0% |
### Coding
| Benchmark | GLM-5.2 | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Pro | 62.1% | 69.2% | 58.6% | 54.2% |
| Terminal-Bench 2.1 | 81.0% | 85.0% | 84.0% | 74.0% |
| ProgramBench | 63.7% | 71.9% | 70.8% | 39.5% |
| NL2Repo | 48.9% | 69.7% | 50.7% | 33.4% |
| DeepSWE | 46.2% | 58.0% | 70.0% | 10.0% |
### Long-Horizon Agentic
| Benchmark | GLM-5.2 | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| FrontierSWE | 74.4% | 75.1% | 72.6% | 39.6% |
| PostTrainBench | 34.3% | 37.2% | 28.4% | 21.6% |
| SWE-Marathon | 13.0% | 26.0% | 12.0% | 4.0% |
| MCP-Atlas | 76.8% | 77.8% | 75.3% | 69.2% |
| Tool-Decathlon | 48.2% | 59.9% | 55.6% | 48.8% |
Key takeaways: GLM-5.2 leads all open-weights models on every agentic benchmark. On FrontierSWE it trails Opus 4.8 by only 1% while beating GPT-5.5. On PostTrainBench it outperforms both GPT-5.5 and Gemini. On AIME 2026 it achieves 99.2%, the highest score among all models listed. The model's weakest area relative to frontier closed models is SWE-Marathon (ultra-long-horizon tasks) where it trails Opus 4.8 by 13%.
詳細比較
### GLM-5.2 vs Claude Opus 4.8
Claude Opus 4.8 remains the strongest model for the hardest agentic coding tasks, leading on SWE-bench Pro (69.2 vs 62.1), NL2Repo (69.7 vs 48.9), and SWE-Marathon (26.0 vs 13.0). However, GLM-5.2 is competitive on FrontierSWE (74.4 vs 75.1), PostTrainBench (34.3 vs 37.2), and MCP-Atlas (76.8 vs 77.8). The pricing gap is enormous: GLM-5.2's output tokens cost 5.7× less ($4.40 vs $25.00 per 1M). For teams running high-volume coding agents, GLM-5.2 offers compelling cost-performance tradeoffs while the MIT license enables self-hosting.
### GLM-5.2 vs GPT-5.5
GLM-5.2 outperforms GPT-5.5 on most coding benchmarks: SWE-bench Pro (62.1 vs 58.6), FrontierSWE (74.4 vs 72.6), Terminal-Bench (81.0 vs 84.0 in one harness, 82.7 vs 83.4 in another). GPT-5.5 edges ahead on DeepSWE (70.0 vs 46.2) and leads on HLE text-only (41.4 vs 40.5). The cost differential is the biggest factor: GLM-5.2 output is ~6.8× cheaper. Both share 1M context windows. GPT-5.5 benefits from deeper ecosystem integration with OpenAI SDKs, Responses API, and enterprise tooling.
### GLM-5.2 vs Gemini 3.1 Pro
GLM-5.2 significantly outperforms Gemini 3.1 Pro on agentic and coding tasks: FrontierSWE (74.4 vs 39.6), PostTrainBench (34.3 vs 21.6), Terminal-Bench (81.0 vs 74.0). Gemini maintains an edge on knowledge (GPQA 94.3 vs 91.2) and multimodal capabilities. GLM-5.2 is cheaper on output tokens ($4.40 vs $12.00) and offers open-weights deployment. Gemini's strength lies in multimodal reasoning and tighter Google ecosystem integration.
コミュニティ評価
Community reaction has been overwhelmingly positive, with many calling it the 'DeepSeek moment for agentic models.' Lambda Labs reported that 'experienced labs and industry leaders were replacing much of their workloads with GLM' within weeks of release. alphaXiv noted that GLM-5.2 is 'the first open weights model we've tried on our autoresearch pipeline that's proven capable for real research tasks.'
On HuggingFace, the model accumulated 126 upvotes on its launch blog post with comments like 'Super impressive model and release, excited to try an Opus-level open model in open coding agents.' The model card shows 513,061 downloads in the past month and has spawned 15 finetunes, 108 quantizations, and 100 Spaces.
Reinvently's independent 28-task evaluation found GLM-5.2 passed 26 of 28 tasks versus 28 for GPT-5.5 and 22 for Claude Fable 5, at roughly one-eighth the cost of GPT-5.5. The notable finding was that GLM-5.2 refused zero tasks (including correctly handling 3 of 4 jailbreaks) while Claude Fable 5 refused 9 tasks including 5 benign ones.
The local LLM community on SpecPicks reported that the small GLM-5.2 variant runs comfortably on a 12GB RTX 3060 at 30-45 tok/s in Q4 quantization, though the flagship 753B model requires over 1TB VRAM for unquantized serving. Developers are increasingly using GLM-5.2 as a 'subagent model' in multi-tier agent architectures—replacing closed-source models for delegated engineering tasks while keeping frontier closed models as the primary driver.
ユースケース
### 1. High-Volume Coding Agent Automation
For teams running autonomous coding agents on repositories, GLM-5.2's 1M context window and competitive SWE-bench scores make it the most cost-effective option. At $4.40 per million output tokens versus $25-30 for closed alternatives, a team processing 50M tokens/month saves $600-730/month. The MIT license enables running the entire stack on private infrastructure. Best for: startups and engineering teams automating code review, PR generation, and repository-level refactoring at scale.
### 2. Long-Horizon Research and Analysis
GLM-5.2's training specifically for long-horizon tasks, combined with anti-hack mechanisms, makes it suited for extended autonomous research workflows. On PostTrainBench (fine-tuning small models via post-training) it outperforms GPT-5.5 and Gemini. AlphaXiv's adoption for paper reproduction and ablation demonstrates this capability. Best for: research labs, automated literature review, and scientific reproducibility pipelines.
### 3. Cost-Aware Multi-Model Agent Routing
GLM-5.2 excels as a 'subagent model' in tiered architectures where Claude Opus 4.8 or GPT-5.5 serves as the driver. It handles delegated coding tasks, implementation, and verification at 6-7× lower cost. EvoLink and Lambda Labs recommend this pattern. Best for: enterprises already using frontier closed models seeking to reduce costs on routine subtasks without sacrificing quality.
### 4. Self-Hosted Enterprise AI for Regulated Industries
The MIT license and self-hosting capability make GLM-5.2 uniquely positioned for organizations with data sovereignty requirements, regulated codebases, or air-gapped environments. It supports deployment on vLLM, SGLang, transformers, KTransformers, and Ascend NPUs. Best for: financial services, healthcare, government, and defense applications requiring on-premises AI with frontier capabilities.
最新ニュース
GLM-5.2 was released on June 16, 2026, under an MIT open-source license. Key developments include:
• **GLM Coding Plan rollout**: All subscribers received GLM-5.2 access immediately, with peak-hour usage at 3× quota and off-peak at 2× (1× promotional rate through September 2026).
• **ZCode desktop agent**: Z.ai launched ZCode, a desktop coding agent powered by GLM-5.2 with /goal command for long-horizon tasks, SSH remote development, and mobile control capabilities.
• **Effort level control**: GLM-5.2 introduces 'High' and 'Max' thinking effort levels, allowing users to balance capability against latency and computational cost. The Max level provides additional computation for challenging tasks.
• **IndexShare architecture innovation**: The novel architecture reuses indexers across every four sparse attention layers, reducing per-token FLOPs by 2.9× at 1M context while maintaining performance.
• **Anti-hack RL training**: A novel two-stage detection system (rule-based filter + LLM judge) prevents reward hacking during agentic RL training, improving model reliability on real-world coding tasks.
• **Framework support expansion**: GLM-5.2 launched with support for SGLang, vLLM, transformers, KTransformers, Unsloth, and Ascend NPU platforms.
• **U.S. release context**: GLM-5.2 was released one day after the U.S. government restricted access to certain Anthropic models, positioning it as a viable alternative for developers affected by access limitations.