GLM-5.2 review: open weights, a 1M context and SWE-bench Pro 62.1
GLM-5.2 (glm-5.2) is the first model to try when you want long-running coding and agent work on open weights at a lower price. The context window is five times GLM-5's, and on coding benchmarks it is close to the closed frontier models.
What changed from GLM-5
Z.ai (Zhipu AI) released GLM-5.2 in June 2026. The headline change from GLM-5 is the context window, up from 200K tokens to about 1M (1,048,576). Maximum output is 131K tokens.
The weights are still MIT licensed: commercial use, modification and fine-tuning are all allowed, about as permissive as model licenses get.
It is text in, text out. For images or audio, look at a multimodal model such as MiniMax M3.
Benchmarks: strongest on long tasks
GLM-5.2 stands out on long software-engineering tasks rather than one-shot problems. On reported figures it scores 62.1 on SWE-bench Pro, ahead of GPT-5.5 (58.6) and its predecessor GLM-5.1 (58.4).
| Model | Weights | SWE-bench Pro |
|---|---|---|
| GLM-5.2 | Open (MIT) | 62.1 |
| GPT-5.5 | Closed | 58.6 |
| GLM-5.1 | Open (MIT) | 58.4 |
It also places near the top of long-horizon evaluations such as FrontierSWE, PostTrainBench and SWE-Marathon, and on FrontierSWE it was reported within about a point of Claude Opus 4.8. The broad consensus is that it is the strongest open-weight model overall as of its release.
Cost
Price is the usual reason to pick an open model. Reports put GLM-5.2 well below GPT-5.5 on cost for equal or better coding results.
Treat that with care. Long agent runs consume a lot of tokens in total, so a low per-token rate can still add up to a large bill. On FastMetal GLM-5.2 is billed per million input and output tokens; current prices are on the model page. The balance is prepaid, so you can cap what an experiment can spend.
Non-English use
GLM-5.2 is multilingual and gives usable answers on general tasks in other languages, including Japanese. We found no language-specific benchmarks for it, though. If register, nuance or domain vocabulary matters, test it on your own prompts rather than going by leaderboard rank.
Good fit, poor fit
Good fit
- Long coding and refactoring jobs
- Multi-step automation with tool calls
- Loading a whole repository into a million-token context
- Commercial projects that need a permissive license
Poor fit
- Images or audio (it is text only)
- Work where quality in a particular language comes first and you cannot spend time testing
- Real-time uses where short, fast replies matter most
FAQ
Is it worth moving from GLM-5?
If you work with long inputs, the jump from 200K to 1M tokens is reason enough. The coding benchmarks improved too.
Can I use GLM-5.2 commercially?
Yes. The weights are MIT licensed, which allows commercial use, modification and fine-tuning.
How good is it outside English?
Usable for general tasks, but we found no language-specific benchmarks. Test it on your own prompts before relying on it.
Try it on FastMetal
GLM-5.2 reasons before it answers, so give it room in max_tokens.
curl https://api.fastmetal.ai/v1/chat/completions \
-H "Authorization: Bearer $FASTMETAL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5.2",
"max_tokens": 16000,
"messages": [{"role": "user", "content": "Refactor this function to remove the duplicated error handling: ..."}]
}'
On FastMetal you can switch between glm-5.2 and closed models like Claude with the same API key and balance. Change model to compare output and cost without touching the rest of your code. The full list is in the model catalog.