# Claude Sonnet 4.6 on the Cotera agent benchmark

> 5/5 every time, billed like a senior engineer. Worth it when correctness is non-negotiable; brutal at scale.

Source: https://cotera.co/benchmarks/models/claude-sonnet-4-6

---

Anthropic's Sonnet 4.6 — the mid-tier in the Claude 4.X family, sitting between Haiku and Opus. The Anthropic crowd's default daily driver for production agent work.

**Score:** 5/5 benchmarks passed · **Cost across the matrix:** $3.83

## Strengths
- 5/5 with the most thorough per-task reasoning of the mid-tier models. Output JSON is consistently the most explicit (full field names, no abbreviation).
- Cleanest Coding answer in the matrix. Wrote the Stripe webhook with HMAC + timing-safe compare, explanation field hit 3 sentences, code field passed every rubric check.
- Zero schema drift across the run. No prose preambles, no envelope leakage, no half-finished JSON.

## Watch-outs
- Marketing cost $1.82 on a single benchmark — 38x Mistral's spend for the same output. Long-context Reddit threads are death by token at Sonnet's rates.
- Total spend was $3.83 — the second-most-expensive 5/5 in the matrix. Unless you've measured a real quality lift on your workload, every Sonnet pass here is a more-expensive version of a Mistral pass.
- Reasoning-tax shows up on retrieval-heavy tasks. CX was $0.723, vs Mistral's $0.066, on the same correct answer.