An open-weight model just outscored GPT-5.5 on SWE-bench Pro — the benchmark closest to what coding agents actually do in production. Z.AI’s GLM-5.2, released June 13 under an MIT license, hits 62.1 on SWE-bench Pro versus GPT-5.5’s 58.6. It runs on a genuine 1-million-token context window, costs $1.40 per million input tokens (versus roughly $8 for GPT-5.5), and the weights are live on HuggingFace today. This is not a “promising open-source alternative.” It is a better model for most coding tasks at a fraction of the price.
The Benchmark Numbers #
GLM-5.2 leads GPT-5.5 across three of the most meaningful coding evaluations available:
SWE-bench Pro: 62.1 vs GPT-5.5’s 58.6. This is the benchmark that measures fixing real GitHub issues in production codebases — not contrived puzzles.FrontierSWE: 74.4% vs GPT-5.5’s 72.6%. Long-horizon tasks simulating multi-step agent work. GLM-5.2 sits within 0.7 percentage points of Claude Opus 4.8 (75.1%).Terminal-Bench 2.1: 81.0 — four points behind Opus 4.8 (85.0) but clearly ahead of GPT-5.5.Design Arena Code:#1 by human preference vote, 10 Elo points above Claude Fable 5. Real developers preferred its output in head-to-head comparisons.
Z.AI launched GLM-5.2 without publishing these numbers themselves — they let third-party evaluators run the tests. That is a confident move, and the results justified it. Independent scores are tracked at BenchLM.ai.
The Cost Math Is Not Close #
If you are running a production coding agent on GPT-5.5 today, GLM-5.2 is worth a serious look. Here is the direct comparison:
| Model | SWE-bench Pro | Input (per 1M tokens) | Output (per 1M tokens) | License |
|---|---|---|---|---|
| GLM-5.2 | 62.1 | $1.40 | $4.40 | MIT |
| GPT-5.5 | 58.6 | ~$8.00 | ~$25.00 | Proprietary |
| Claude Opus 4.8 | ~63 | ~$15.00 | ~$75.00 | Proprietary |
A team spending $25,000 per month on GPT-5.5 for a coding pipeline could run the same workload on GLM-5.2 for approximately $4,000. GLM-5.2 also supports prompt caching, dropping the effective cached input cost to $0.26 per million tokens — which matters in agent loops that re-read the same context repeatedly. VentureBeat’s full cost breakdown covers additional provider comparisons.
What MIT License Actually Means Here #
Most “open” AI models are open in name only. GLM-5.2 is MIT-licensed: fine-tune it, run it commercially, redistribute derivatives — and no one can revoke your access. The weights are at huggingface.co/zai-org/GLM-5.2 with no waiting list or application process.
Compare this to DeepSeek, which carries commercial restrictions that disqualify it for many enterprise workloads. GLM-5.2’s MIT license is a genuine differentiator in this tier of open-weight models.
Local deployment requires 256GB of unified memory for the 2-bit GGUF quantization, which puts it out of reach for most individual setups. The API is the practical path for teams.
The 1M Context Window Is Real #
GLM-5.2’s 1M-token context is enabled by IndexShare — a sparse attention mechanism that shares an attention index across every four transformer layers, cutting per-token FLOPs by 2.9x at full context length. This is not a marketing claim with degraded performance at scale; the architecture is built for it.
The practical implication: a coding agent can hold an entire mid-sized repository, its full task transcript, and the relevant documentation in a single context window. No chunking. No retrieval-augmented workarounds. GLM-5.1 (the predecessor) sustained approximately 1,700 agent steps in one session and ran autonomous loops for up to eight hours. GLM-5.2 extends that further.
How to Start Using It #
The fastest path is Ollama:
ollama run glm-5.2:cloud
This routes through Z.AI’s infrastructure with the Ollama interface — no local hardware required. For production use, the Z.AI API is OpenAI-compatible, so existing integrations need minimal changes:
from openai import OpenAI
client = OpenAI(
base_url="https://open.bigmodel.cn/api/paas/v4/",
api_key="YOUR_KEY"
)
response = client.chat.completions.create(
model="glm-5.2",
messages=[{"role": "user", "content": "Review and refactor this module..."}]
)
OpenRouter ($0.95/$3.00 per million tokens) and Together AI offer third-party hosting if you prefer not to use Z.AI directly.
The Bottom Line #
The open-source versus closed-source AI debate has mostly been philosophical. GLM-5.2 makes it financial. Better SWE-bench Pro scores than GPT-5.5, an MIT license, genuine 1M-token context, and a price that is 6x lower. If you are building coding agents or long-horizon pipelines, the burden of proof has shifted: you now need a reason not to evaluate GLM-5.2 before committing to a proprietary alternative.