# GLM 5.2 on the Cotera agent benchmark

> The cheap open-weights contender — 3/5 for $0.52, tripped up only by JSON hygiene.

Source: https://cotera.co/benchmarks/models/glm-5.2

---

Zhipu AI's GLM 5.2, run through OpenRouter. A 1M-context open-weights model from the Beijing lab; the successor to the GLM 5.1 that also appears in this matrix.

**Score:** 3/5 benchmarks passed · **Cost across the matrix:** $0.521

## Strengths
- Cheapest way to clear three benchmarks: $0.52 total.
- Solid tool use — clean passes on Crunchbase funding, CX complaint mining, and Apollo pricing extraction, with tight 2-4 tool-call loops.
- Native tool-calling worked out of the box through OpenRouter; no text-format tool-call leakage like the Nemotron family.

## Watch-outs
- Final-JSON discipline is the weak spot. On the coding task it truncated the answer mid-string — the `code` field was cut off before the closing brace, so the object never parsed.
- On the Reddit task it produced a final message that simply wasn't valid JSON.
- Both failures were about *emitting* clean, complete JSON — not about gathering the right data, which it generally did.