GLM-5.1: First Open-Source Model to Top SWE-bench Pro (2026)

On April 7, 2026, Z.ai dropped GLM-5.1 — a 754-billion-parameter open-weight model that, for nine days, held the top spot on SWE-bench Pro with a score of 58.4%, beating GPT-5.4 (57.7%) and Claude Opus 4.6 (57.3%). It was the first time an open-source model had ever led that leaderboard. The weights are MIT-licensed. You can download them, run them locally, or call them via a cheap OpenAI-compatible API. If you work with code — as a developer, a team lead, or someone building AI-assisted workflows — GLM-5.1 is the most significant open-source release of 2026 so far.

This guide covers everything: the benchmarks, the architecture, how to access the model via API and locally, how to plug it into coding agents like Claude Code and Cline, pricing versus GPT-5.4 and Opus 4.7, and who should actually switch. The model is live. Here is how to use it.

What Is Z.ai? A Quick Background

Z.ai is the new public name for Zhipu AI — a Chinese AI lab spun out of Tsinghua University that has been building foundation models since 2019. In January 2026, the company became the first publicly traded AI foundation model company in the world, raising approximately $558 million on the Hong Kong Stock Exchange. The GLM (General Language Model) series has been its flagship, tracking closely behind the frontier proprietary models for the past two years. GLM-5.1 is where that trajectory crossed a meaningful line.

The model was trained entirely on 100,000 Huawei Ascend 910B chips — no US silicon, no NVIDIA H100s or H200s. This is not just a geopolitical data point. It is evidence that the Ascend ecosystem has reached a threshold where it can produce genuinely frontier-competitive results. The infrastructure independence matters for organizations concerned about supply chain risk, and it means Z.ai’s training capacity is not constrained by US export controls.

The Benchmark Breakdown

Here is what GLM-5.1 actually scored and what those numbers mean in practice.

SWE-bench Pro: The Defining Metric

SWE-bench Pro tests whether an AI model can autonomously resolve real GitHub issues on production codebases — the kind of work software engineers do every day. It is widely considered the most meaningful benchmark for agentic coding capability because it requires multi-step reasoning, file navigation, code editing, and verification, not just pattern matching on training data.

• GLM-5.1: 58.4% (SOTA at release, April 7, 2026)
• GPT-5.4: 57.7%
• Claude Opus 4.6: 57.3%
• Claude Opus 4.7: 64.3% (current leader as of April 16, 2026)

GLM-5.1 held the top position for nine days before Claude Opus 4.7 displaced it. For a free, self-hostable open-weight model to beat every proprietary offering for any period is a historic milestone. For context: twelve months ago, the gap between open-source models and frontier proprietary models on SWE-bench was approximately 20 percentage points. That gap has closed to under 6.

Code Arena and CyberGym

Beyond SWE-bench, GLM-5.1 scores 1,530 on Code Arena Elo — a live, human-preference ranking of coding model outputs. It is the first open-weight model to enter the top three on that leaderboard. On CyberGym, a security task completion benchmark across 1,507 real-world tasks, GLM-5.1 leads all models at 68.7%, suggesting strong performance on security analysis and vulnerability reasoning.

Where GLM-5.1 does not lead: advanced mathematical reasoning. On AIME 2026, GPT-5.4 scores 98.7% and Claude Opus 4.6 scores 98.2%, while GLM-5.1 comes in at 95.3%. For pure math competition problems, frontier proprietary models retain a meaningful edge. For software engineering work, GLM-5.1 is now firmly in the same tier.

What Makes GLM-5.1 Different: The Long-Horizon Architecture

Z.ai describes GLM-5.1 as designed for “long-horizon agentic tasks” — work that requires maintaining context and coherence across hours of autonomous execution, not just a single code edit. The model supports sustained 8-hour autonomous execution windows, which is directly relevant for production coding agents handling multi-file refactors, end-to-end feature implementations, or complex debugging sessions without human checkpoints.

Key architectural capabilities:

• 754 billion parameters at full precision, with quantized variants for local deployment
• Long context window: supports extended token sequences for full codebase context
• Native function calling: tool use is first-class, not retrofitted
• Thinking mode: extended reasoning chains for complex multi-step problems
• Structured outputs: JSON mode and schema-constrained generation
• Context caching: significantly reduces costs on repeated similar queries

The OpenAI-compatible API surface means you can drop GLM-5.1 into any existing integration without changing your code — just swap the base URL and model name.

How to Access GLM-5.1

Option 1: Z.ai API (Fastest Setup)

The simplest path is the Z.ai managed API, which is OpenAI-compatible. You can be running it in under two minutes:

1. Create an account at z.ai and generate an API key from the dashboard
2. Point your existing OpenAI client to Z.ai’s base URL: https://api.z.ai/api/paas/v4/
3. Set the model name to glm-5.1

In Python:

from openai import OpenAI

client = OpenAI(
    api_key="your-z-ai-api-key",
    base_url="https://api.z.ai/api/paas/v4/"
)

response = client.chat.completions.create(
    model="glm-5.1",
    messages=[{"role": "user", "content": "Refactor this function to handle edge cases."}]
)
print(response.choices[0].message.content)

The same pattern works with any OpenAI-compatible client: the TypeScript SDK, LangChain, LlamaIndex, or a raw HTTP call. You do not need any Z.ai-specific library.

Option 2: OpenRouter

GLM-5.1 is available on OpenRouter as z-ai/glm-5.1, which gives you a unified API key across multiple models. Useful if you are already using OpenRouter for multi-model routing or do not want another vendor account to manage.

Using GLM-5.1 With Coding Agents

Z.ai has explicitly designed GLM-5.1 to work as the intelligence layer behind popular coding agents. Supported integrations include Claude Code, OpenCode, Kilo Code, Roo Code, Cline, and Droid — the primary agentic coding tools used by professional development teams in 2026.

Claude Code Integration

To route Claude Code sessions through GLM-5.1, add these environment variables to your shell profile or Claude Code configuration. Z.ai provides an Anthropic-compatible adapter endpoint that translates the Claude API format to GLM-5.1:

export ANTHROPIC_DEFAULT_SONNET_MODEL="glm-5.1"
export ANTHROPIC_DEFAULT_OPUS_MODEL="glm-5.1"

With these set, Claude Code routes its model calls through Z.ai using GLM-5.1 as the backend. This is the fastest way to benchmark GLM-5.1’s agentic coding performance on your actual codebase without changing your editor or workflow. Session length and tool-use patterns remain identical to a standard Claude Code session — only the underlying model changes.

Cline, Roo Code, and Kilo Code

For VS Code-based agents like Cline, Roo Code, and Kilo Code, configure the model provider in the extension settings:

• Set Provider to “OpenAI Compatible”
• Set Base URL to https://api.z.ai/api/paas/v4/
• Set API Key to your Z.ai key
• Set Model to glm-5.1

All three agents support this configuration path natively. GLM-5.1’s native function-calling capability means tool use — file reads, shell commands, browser calls — works reliably without special prompt engineering.

Local Deployment: vLLM and Ollama

For teams with data residency requirements or high-volume workloads where self-hosting is more economical, GLM-5.1 weights are available on Hugging Face at zai-org/GLM-5.1 under the MIT license. Two inference paths are production-ready:

vLLM (Recommended for Production)

vLLM is the standard choice for high-throughput production deployments. Serving GLM-5.1 follows the standard vLLM pattern:

vllm serve zai-org/GLM-5.1   --tensor-parallel-size 8   --max-model-len 32768

Hardware requirements at full precision are substantial: 754 billion parameters require significant VRAM spread across multiple GPUs. Z.ai provides GPTQ and AWQ quantized variants that reduce memory requirements to a range accessible on smaller multi-GPU setups. The quantized models see approximately 2-3 percentage points of SWE-bench degradation, which still places them above GPT-5.4 and Claude Opus 4.6.

Ollama (Recommended for Development)

For local development and evaluation, Ollama provides the simplest deployment path. Check the zai-org/GLM-5.1 Hugging Face repository for the latest Ollama-compatible model files. Local deployment is practical for teams doing workflow evaluation or testing privately before committing to cloud inference costs at scale.

Pricing: GLM-5.1 vs. GPT-5.4 vs. Claude Opus 4.7

Cost is where GLM-5.1 is unambiguously ahead. Z.ai prices the model at approximately $0.95 per million input tokens and $3.15 per million output tokens, with cached inputs at $0.26 per million. Context caching matters significantly for coding agents that repeatedly load large codebases into context across long sessions.

Compared to frontier proprietary models:

• GLM-5.1 (Z.ai API): ~$0.95 input / $3.15 output per million tokens
• GPT-5.4 (OpenAI): ~$5.00 input / $15.00 output per million tokens
• Claude Opus 4.7 (Anthropic): ~$7.50 input / $24.00 output per million tokens

At current pricing, GLM-5.1 is roughly 5x cheaper on input tokens and 5-8x cheaper on output tokens compared to the leading proprietary frontier models. For high-volume coding agent workloads where a single agentic session consumes millions of tokens, this cost difference is the deciding factor for many teams. An 8-hour autonomous coding session consuming 10 million output tokens costs approximately $240 at Opus 4.7 pricing and about $31 at GLM-5.1 pricing. That is not a marginal difference; it is a budget category change.

Who Should Make the Switch?

The right candidates for GLM-5.1 are clear:

• Teams running high-volume agentic coding pipelines: If your team runs Cline, Roo Code, or similar agents for multiple developers, the per-token savings accumulate into meaningful budget relief within weeks.
• Organizations with data residency requirements: Self-hosted GLM-5.1 means code never leaves your environment. The MIT license removes any legal ambiguity around deployment or fine-tuning.
• Security and vulnerability research teams: GLM-5.1’s #1 ranking on CyberGym suggests specific strength on security reasoning. Teams doing defensive security work may find it outperforms frontier models on domain-specific tasks.
• Developers evaluating open-source model quality: If you have assumed open-source models are categorically behind proprietary frontier models, GLM-5.1 is the most persuasive counterexample to date. Running it on your own codebases is now a practical exercise, not an academic one.

Who should not switch without testing: teams where mathematical reasoning is a primary workload, or where Claude Opus 4.7’s current SWE-bench lead translates to meaningfully better output on your specific tasks. Run your own evals on representative samples before making a production change. Aggregate benchmarks are the starting point for evaluation, not the ending point.

Conclusion

GLM-5.1 is a genuine milestone. A free, MIT-licensed, self-hostable model that held the top spot on the most rigorous coding benchmark for nine days — beating GPT-5.4 and Claude Opus 4.6 — represents a structural shift in what open-source AI can deliver. The cost advantage over proprietary models is not marginal; it is 5-8x. The training-on-Ascend story matters beyond geopolitics: it demonstrates that frontier AI quality is no longer exclusive to NVIDIA-based clusters.

Claude Opus 4.7 currently leads on SWE-bench Pro at 64.3%, and proprietary models retain an advantage in pure mathematical reasoning. But the gap between the best open-source and best proprietary coding models has collapsed from roughly 20 percentage points to under 6 in twelve months. At this trajectory, the question in 2027 will not be whether open-source models can compete, but which open-source model your team runs.

For developers and teams building AI-assisted engineering workflows today: GLM-5.1 is ready for production evaluation. Download the weights, call the API, or plug it into your coding agent. The model earned its place in the frontier tier — and at a fraction of the cost.