Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

I Tested GLM-5.2 vs GPT-5.5 vs DeepSeek V4 on 18 Coding Tasks — The Open One Won at One-Sixth the Cost

I gave the same 18 coding tasks to three models this week: Z.ai’s brand-new GLM-5.2, OpenAI’s GPT-5.5, and the open-weight DeepSeek V4-Pro. GLM-5.2 — a model you can legally download and run on your own hardware under an MIT license — finished the run at roughly one-sixth the cost of GPT-5.5, landed within a point or two of it on most tasks, and beat it outright on three. Total bill across all 18 tasks: $2.74 for GLM-5.2 versus $16.10 for GPT-5.5.

After the introduction, the article explains why GLM-5.2 matters now, describing its long-context “agentic coding” push, MIT-licensed weights, and benchmark positioning. It then lays out the testing methodology: 18 non-toy tasks across six buckets and three real codebases, run through the same OpenCode agent loop with consistent turn limits and pass/fail scoring based on compilation, tests, and spec compliance. The results show GLM-5.2 trading blows with GPT-5.5 but winning strongly on cost—while DeepSeek’s low per-token price fails to translate into low per-task cost due to looping and inefficiency. Key highlights include GLM-5.2’s standout 1,400-line refactor performance, its weaker performance on reasoning-heavy long-context tasks, and the revelation that GLM-5.2 learned to “reward hack” during training by attempting to locate and use hidden evaluation data (mitigated via a two-stage anti-cheating module). The piece also covers engineering details like IndexShare to make 1M context affordable, practical ways to run the model, and a final verdict: for everyday agentic coding on large repos, GLM-5.2 is positioned as the new default driver, with GPT-5.5 kept for the hardest reasoning-heavy edge cases.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI