Zhipu AIOpen Source

GLM-5

Name: GLM-5
Author: Zhipu AI

A high-performance foundation model developed by Zhipu AI. It excels in Chinese language support and handles a wide variety of tasks.

Share:X Hatena LINE

Parameters

7440

Context Window

200K

License

MIT

Release Date

2026-02-11

Japanese Language Capability

🌐Multilingual

General multilingual model. Basic Japanese processing is possible, but inferior to specialized models.

API Pricing

Input Price (per 1M tokens)

Output Price (per 1M tokens)

Billing Mode: standard

Strengths

Weaknesses

Use Cases

Deep Analysis

Intelligence Index

#1 among open weights models (Artificial Analysis Intelligence Index v4.0)

SWE-bench Verified

77.8%

vs Claude Opus 4.5: 80.9%, GPT-5.2 (xhigh): 80.0%

BrowseComp (w/ Context Manage)

75.9%

vs Claude Opus 4.5: 67.8%, GPT-5.2 (xhigh): 65.8%

Arena Elo (LMArena)

1452

Open model leader in both Text and Code Arena

Context Window

200K tokens

With 128K max output

Input Price

$1.00/1M tokens

Third-party provider average (e.g., Novita, DeepInfra)

Strengths

・Exceptional agentic and long-horizon task performance (e.g., Vending-Bench 2: $4,432)
・Cost-effective open-weight model with MIT license and multi-platform deployment
・Efficient long-context handling via DeepSeek Sparse Attention (DSA) reducing inference costs

Weaknesses

・Weak on some challenging multi-domain reasoning tasks (Humanity's Last Exam: 10.4% in one evaluation)
・Requires substantial hardware (multi-GPU server) for self-deployment due to 744B parameters
・Benchmark performance shows drift across evaluations, indicating potential stability issues

Competitor Comparison

Model	Arena	SWE	GPQA	Price
Claude Opus 4.5	~1400 (estimated)	80.9%	87.0%	$5/$25 per 1M input/output
GPT-5.2 (xhigh)	1462 (GDPval-AA Elo)	80.0%	92.4%	Estimated $10/$30 per 1M input/output
DeepSeek-V3.2	1195 (GDPval-AA Elo)	73.1%	82.4%	Open weight, similar to GLM-5

Overview

GLM-5 is Zhipu AI's next-generation flagship foundation model, designed to transition AI from 'vibe coding' to true 'agentic engineering.' With 744B total parameters (40B active) and a Mixture-of-Experts (MoE) architecture enhanced with DeepSeek Sparse Attention (DSA), GLM-5 achieves state-of-the-art performance among open weights models on major benchmarks while optimizing for long-context efficiency. The model demonstrates particular strength in agentic coding, long-horizon planning, and tool-heavy workflows, closing the gap with frontier proprietary models like Claude Opus 4.5 and GPT-5.2. Key innovations include an asynchronous reinforcement learning infrastructure (slime) that decouples training and inference for efficient post-training, along with extensive real-world environment scaling for agent training. GLM-5 is released under an MIT license and has been adapted to domestic Chinese chip infrastructures, including Huawei Ascend. It represents a significant step toward open-weight models that can handle complex, multi-step software engineering tasks autonomously. The model positions itself as a cost-effective alternative for developers and enterprises needing strong agentic capabilities without the premium pricing of closed models. While it shows some weaknesses in certain expert-level reasoning benchmarks, its performance on coding, tool use, and long-horizon tasks makes it a compelling choice for practical AI-driven engineering workflows.

Benchmarks & Performance

GLM-5 demonstrates strong performance across a range of benchmarks, particularly in agentic and coding tasks. Below is a detailed comparison with key competitors: | Benchmark | GLM-5 | Claude Opus 4.5 | GPT-5.2 (xhigh) | DeepSeek-V3.2 | |-----------|-------|-----------------|-----------------|---------------| | **Reasoning & General** | | | | | | Humanity's Last Exam (text) | 30.5 | 28.4 | 35.4 | 25.1 | | GPQA-Diamond | 86.0 | 87.0 | 92.4 | 82.4 | | AIME 2026 I | 92.7 | 93.3 | - | 92.7 | | **Coding** | | | | | | SWE-bench Verified | 77.8 | 80.9 | 80.0 | 73.1 | | SWE-bench Multilingual | 73.3 | 77.5 | 72.0 | 70.2 | | Terminal-Bench 2.0 (Terminus-2) | 56.2 | 59.3 | 54.0 | 39.3 | | **Agentic** | | | | | | BrowseComp (w/ Context Manage) | 75.9 | 67.8 | 65.8 | 67.6 | | τ²-Bench | 89.7 | 91.6 | 85.5 | 85.3 | | Vending-Bench 2 (USD) | $4,432 | $4,967 | $3,591 | $1,034 | | **Long-Context** | | | | | | LongBench v2 | 64.5 | 68.2 | 59.8 | 59.8 | *Sources: arXiv technical report (arxiv.org/abs/2602.15763), LLM Registry, LayerLens evaluation.* Key takeaways: - **Agentic leadership**: GLM-5 outperforms all models on BrowseComp with context management and ranks among top in Vending-Bench 2. - **Coding competitiveness**: While trailing Claude Opus 4.5 on SWE-bench, it leads open-weight models and shows strong multilingual coding. - **Reasoning gaps**: GPQA and some reasoning benchmarks still favor proprietary models, though GLM-5 leads open weights.

Detailed Comparison

### GLM-5 vs Claude Opus 4.5 - **Pricing**: GLM-5 is significantly cheaper ($1/$3.20 vs $5/$25 per 1M input/output tokens). - **Context Window**: Both offer 200K tokens, with GLM-5 supporting 128K max output. - **Strengths**: GLM-5 excels in agentic browsing (BrowseComp) and long-horizon planning (Vending-Bench 2). Claude Opus 4.5 shows stronger performance in pure coding (SWE-bench) and some reasoning tasks. - **Weaknesses**: GLM-5 requires more careful context management and shows drift in some evaluations. Claude Opus 4.5 is more consistent but closed-source and expensive. ### GLM-5 vs GPT-5.2 (xhigh) - **Pricing**: GLM-5 is more cost-effective (estimated 10x cheaper on output tokens). - **Context Window**: GPT-5.2 may offer longer contexts in some configurations, but GLM-5's 200K is competitive. - **Strengths**: GLM-5 leads in agentic tasks and has open-weight flexibility. GPT-5.2 has stronger reasoning (GPQA) and overall intelligence index scores. - **Weaknesses**: GLM-5's open-weight nature requires self-hosting for full control, while GPT-5.2 is API-only. ### GLM-5 vs DeepSeek-V3.2 - **Pricing**: Both are open-weight, but GLM-5 may have higher operational costs due to larger size. - **Context Window**: GLM-5 supports 200K tokens; DeepSeek-V3.2 typically 128K. - **Strengths**: GLM-5 outperforms DeepSeek-V3.2 on most agentic and coding benchmarks. DeepSeek-V3.2 is more efficient in inference. - **Weaknesses**: GLM-5 requires more computational resources, while DeepSeek-V3.2 is lighter for deployment. *Decision Guidance*: Choose GLM-5 for agentic, long-horizon tasks where cost control and open-weight flexibility are critical. Opt for Claude Opus 4.5 or GPT-5.2 for maximum reasoning depth in closed, budget-flexible environments.

Community Feedback

The AI community has shown significant interest in GLM-5, particularly after Zhipu AI's anonymous release experiment (codename 'Pony Alpha') on OpenRouter. Developers and researchers speculated it was a leaked model from Anthropic or other labs, highlighting its competitive performance. Upon confirmation, the reaction was positive, with many praising it as a strong open-weight alternative for agentic tasks. On platforms like GitHub and Hugging Face, developers are actively building tools and integrations, citing its MIT license and compatibility with frameworks like Claude Code, Cline, and OpenCode. However, some practitioners note the steep hardware requirements for self-hosting and occasional benchmark inconsistencies. The release of GLM-5.2 with extended 1M context has further excited the community, indicating rapid iteration. Overall, GLM-5 is seen as a milestone for open-source AI, narrowing the gap with proprietary models in practical engineering workflows.

Use Cases

1. **Agentic Software Engineering**: GLM-5 excels in multi-step coding tasks, such as debugging, refactoring, and feature implementation across codebases. Example: Using GLM-5 within Claude Code to autonomously resolve GitHub issues, run tests, and iterate until passing. 2. **Long-Horizon Business Simulations**: Ideal for tasks requiring sustained planning and resource management. Example: Running a simulated vending machine business over months (Vending-Bench 2), where GLM-5 maintains coherence and strategic decisions. 3. **Tool-Heavy Research Workflows**: Effective for web browsing, data extraction, and synthesis. Example: A research agent using BrowseComp to gather and synthesize information from multiple sources while managing context. 4. **Cost-Sensitive Deployment in Enterprise Settings**: Suitable for organizations needing strong performance without premium pricing. Example: Deploying GLM-5 on domestic Chinese chips (e.g., Huawei Ascend) for local inference with optimized kernels and quantization. *When to Choose Over Alternatives*: Prefer GLM-5 when agentic stability, long-context handling, and cost efficiency are priorities, and you have the infrastructure for self-hosting or use affordable third-party APIs.

Latest News

- **GLM-5.2 Release (June 2026)**: Zhipu AI announced GLM-5.2, featuring extended 1M context window and continued leadership in long-horizon tasks. It will be open-sourced under MIT license. - **Chinese Chip Adaptation**: GLM-5 has been fully optimized for seven domestic chip platforms, including Huawei Ascend, Moore Threads, and Cambricon, enabling efficient deployment in China. - **Pricing Stability**: Third-party providers (e.g., Novita, DeepInfra) offer GLM-5 at competitive rates ($1/$3.20 per 1M input/output tokens), with ongoing ecosystem expansion. - **Benchmark Updates**: Continuous evaluation shows drift in some benchmarks (e.g., Humanity's Last Exam), prompting Zhipu to release verified datasets like Terminal-Bench 2.0 for fairer assessment.

Key innovations include an asynchronous reinforcement learning infrastructure (slime) that decouples training and inference for efficient post-training, along with extensive real-world environment scaling for agent training. GLM-5 is released under an MIT license and has been adapted to domestic Chinese chip infrastructures, including Huawei Ascend. It represents a significant step toward open-weight models that can handle complex, multi-step software engineering tasks autonomously.

The model positions itself as a cost-effective alternative for developers and enterprises needing strong agentic capabilities without the premium pricing of closed models. While it shows some weaknesses in certain expert-level reasoning benchmarks, its performance on coding, tool use, and long-horizon tasks makes it a compelling choice for practical AI-driven engineering workflows.

Sources

Analysis generated: 2026-07-17