# Kimi K2.6 on the Cotera agent benchmark

> Confident prose, missing code. Token-repetition loop on Reddit. Read the actual outputs before you trust this one.

Source: https://cotera.co/benchmarks/models/fw-kimi-k2p6

---

Moonshot AI's Kimi K2.6 — a Chinese long-context model with reasoning capabilities. We tested the production endpoint as of November 2026.

**Score:** 3/5 benchmarks passed · **Cost across the matrix:** $1.27

## Strengths
- Sales, CX, and Web Scraping all passed cleanly. When Kimi K2.6 stays in its lane, it's a perfectly competent agent.
- CX pass was the only Kimi run under $0.35 — 3 tool calls, clean grouping of complaints.

## Watch-outs
- Reddit failure was a decoder pathology: the model emitted `{"sourceUrl":"..."` and then got stuck looping the same key (`{"sourceUrl":{"sourceUrl":...`) until it was killed. Token-repetition mid-JSON. Cost $0.423 of nonsense.
- Coding failure was a confidence/competence mismatch: valid JSON, confident explanation describing HMAC-SHA256, but the code field omitted the actual `createHmac` call. Described the algorithm without implementing it.
- Both failure modes are dangerous because they look correct at a glance. The Coding output passes a syntax check; only the rubric's substring check (`createHmac("sha256", ...)`) caught it.