An
claude-opus-4-8 Benchmark & Insights
Anthropic Claude API
Updated Jul 18, 2026 All models
Sample size
156 runs
in window
Accuracy
93.8%
consensus match · 161d
Confidence
82%
over 162 runs
Window end
Jul 18, 2026
most recent run
Input price
$5.00/MTok
prompt tokens
Output price
$25.00/MTok
completion tokens
Model insights
- 01 The strongest Claude, but every one of its 11 misses is an underrating and it said "safe" on three consensus-"unsafe" days (2026-02-18, 02-25, 07-05) — the wrong direction to fail for a safety check.
- 02 At $5/$25 it's an expensive optimist; haiku-4-5 gets within 2.5 points at a fifth of the price.
Notes
Improved version of Opus 4.7. Release notes for Claude Opus 4.8
Recent forecasts