# Claude Sonnet 5 on the Cotera agent benchmark

> The other clean 5/5. Never fumbled a final answer, never over-searched, never mangled JSON.

Source: https://cotera.co/benchmarks/models/claude-sonnet-5

---

Anthropic's latest Sonnet, tested via the first-party Claude API. The mid-tier of the Claude 5 line — below Opus on raw capability but tuned for agentic tool use and instruction-following.

**Score:** 5/5 benchmarks passed · **Cost across the matrix:** $1.44

## Strengths
- Swept 5/5 with the least drama of any model in the suite — every run produced a clean, schema-conforming JSON object on the first try.
- Knows when to stop. Averaged ~2.4 tool calls per benchmark; it gathered what it needed and answered instead of double-checking itself into a loop — the exact failure mode that sank the open-weights models on the Reddit task.
- Passed the marketing-reddit benchmark (rank 5 tools by distinct-commenter mentions) that Nemotron and GLM both looped on and never concluded.

## Watch-outs
- The priciest sweep here at standard Sonnet pricing ($3/$15 per MTok): $1.44 for the matrix — ~5.3x Mistral Large 3 and ~2.8x GLM 5.2, though still a fraction of Opus.
- One operational gotcha: Sonnet 5 rejects the `temperature` parameter (like GPT-5 and the o-series). Send it and the API 400s and you get an empty response — you must drop `temperature`/`top_p` for this model.
- Short-horizon agent tasks only (2-3 tool calls here). No data on long multi-hour chains or adversarial inputs.