By ByteBot
A Chinese open-source model just beat GPT-5.5 on the benchmark most developers actually trust for production coding — and it costs one-sixth as much to run. Zhipu AI’s GLM-5.2, released June 13, scored 62.1 on SWE-bench Pro versus GPT-5.5’s 58.6. It ships under the MIT license, runs on OpenRouter right now at $1.40 per million input tokens, and its full weights are on Hugging Face. If you’re still defaulting to GPT-5.5 for every coding workload, the math just changed.
What GLM-5.2 Is
GLM-5.2 is a 744-billion-parameter mixture-of-experts model from Zhipu AI (operating as Z.ai), a Beijing-based lab. Only 40 billion parameters are active per token — the MoE architecture keeps compute cost manageable while maintaining frontier-level quality. The context window is one million tokens, enabled by IndexShare sparse attention that cuts per-token FLOPs by 2.9x at full context length. It was trained on 28.5 trillion tokens.
Two thinking effort modes are available: High (balanced speed and performance) and Max (full reasoning, up to 85k output tokens per complex task). You can disable thinking entirely with enable_thinking=false for latency-sensitive pipelines. Multi-Token Prediction speculative decoding improves throughput by roughly 20%.
The Benchmark Numbers
The headline result is SWE-bench Pro, which measures a model’s ability to resolve real GitHub issues — production-grade code repair, not toy problems. GLM-5.2 scores 62.1 against GPT-5.5’s 58.6, a meaningful gap. Claude Opus 4.8 still leads at 69.2, but it’s also not close to GLM-5.2 in price.
| Model | SWE-bench Pro | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|---|
| Claude Opus 4.8 | 69.2 | $5.00 | $25.00 |
| GLM-5.2 | 62.1 | $1.40 | $4.40 |
| GPT-5.5 | 58.6 | $5.00 | $30.00 |
GLM-5.2 also took the top spot on BridgeBench Reasoning — beating Fable 5 — and scored 81.0 on Terminal-Bench 2.1 (Claude leads at 85.0, GPT-5.5 at 84.0). On FrontierSWE and MCP-Atlas, GLM-5.2 trails Claude Opus 4.8 by less than one point. It is the only open-weight model currently competing on the Arena agent leaderboard alongside closed frontier models. Per the VentureBeat analysis, GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for one-sixth the cost.
The Cost That Changes Your Model Selection
At GPT-5.5’s output rate of $30 per million tokens, a team running 10M input and 5M output tokens per month pays $200. The same workload on GLM-5.2 costs $36. That gap compounds fast at scale. The model isn’t claiming best-in-class on every task — it’s claiming better than GPT-5.5 on SWE-bench specifically while being dramatically cheaper. That’s enough to force a re-evaluation.
How to Use It Today
The fastest path is OpenRouter, which gives you an OpenAI-compatible endpoint with multiple provider backends routing under the hood:
from openai import OpenAI
client = OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="your-openrouter-key"
)
response = client.chat.completions.create(
model="z-ai/glm-5.2",
messages=[{"role": "user", "content": "Fix this Python function..."}],
extra_body={"thinking": {"type": "enabled", "budget_tokens": 10000}}
)
print(response.choices[0].message.content)For self-hosting, the full weights are at zai-org/GLM-5.2 on Hugging Face (FP8 quantized version also available). vLLM works out of the box:
vllm serve "zai-org/GLM-5.2"MIT license. No usage restrictions. No regional limits. Commercial use is unrestricted.
The Caveat Worth Taking Seriously
Open weights cut both ways. Because anyone can download GLM-5.2, modify it, and remove its safety guardrails, security researchers have already observed discussions on underground forums about fine-tuning it for offensive use. Independent evaluations by Graphistry and Semgrep found it on par with US frontier models for vulnerability discovery — which matters for the threat landscape, not just your stack.
For enterprise users: do not send source code, customer data, vulnerability reports, or export-controlled information through Z.ai’s API without legal review. This applies to regulated sectors (defense, finance, critical infrastructure) more acutely. Self-hosting via Hugging Face sidesteps the data sovereignty question entirely — your tokens never leave your infrastructure.
For personal projects, open-source work, and local development: MIT license, no concerns.
Bottom Line
GLM-5.2 is the most capable open-weight coding model released to date. It beats GPT-5.5 on SWE-bench Pro — the benchmark that best predicts real-world code repair performance — at a fraction of the API cost, or freely self-hosted. Claude Opus 4.8 still leads on the hardest long-horizon tasks, but GLM-5.2 closes that gap to near-nothing on most practical benchmarks while costing dramatically less.
The narrative that China is still catching up on AI is already outdated. On production coding benchmarks, it has caught up. Update your model selection accordingly — or keep paying 6x for GPT-5.5 output that loses on the benchmark that matters.
