# GLM 5.1 on the Cotera agent benchmark

> Smart quotes in Stripe code. Empty final on Reddit. JSON encoding is GLM 5.1's nemesis.

Source: https://cotera.co/benchmarks/models/fw-glm-5p1

---

Zhipu AI's GLM 5.1 — a Chinese reasoning model with open weights. We tested the production endpoint as of November 2026.

**Score:** 3/5 benchmarks passed · **Cost across the matrix:** $1.19

## Strengths
- Sales pass was $0.048 — the second-cheapest Sales pass in the matrix after Gemini 3 Flash.
- CX and Web Scraping passed without drama. When GLM 5.1 commits to writing a final answer, it writes one.

## Watch-outs
- Reddit failure was the worst kind: empty final message after 7 tool calls. The model retrieved everything correctly and then closed the run without writing anything. $0 spent on inference is misleading — you still paid for the tools.
- Stripe failure was a tokenizer/encoding bug: GLM 5.1 emitted the JSON with smart quotes inside the code field and unbalanced backtick template-literal fences. Output was clearly an attempt at the right answer; encoding broke it.
- Both failure modes are decoder-level, not reasoning-level. This is the kind of model where you wrap the output in a strict-JSON repair step before the schema parser sees it.