Artificial Analysis
Artificial Analysis v5.0 Puts Real-World Tasks at the Heart of Model Rankings
Artificial Analysis v5.0 weighs knowledge work, coding, science, and general skills equally, shifting model selection from exam scores to delivery qua
Back to home
Terminal-Bench
3 articles tagged with this topic
Terminal-Bench
Terminal-Bench Flawed in 30% of Tasks; Scores Rise After Repairs
A Stanford audit found 28 defects in 89 Terminal-Bench 2.0 tasks. After fixes, 19 agents gained over 10 points, exposing benchmark maintenance debt.
DeepSeek
DeepSeek V4 Flash Benchmarks Independently Replicated — Numbers Match, But China's LLM Evaluation
Third-party team Ante replicated DeepSeek V4 Flash's 82.7% Terminal-Bench 2.1 score (368/445 successes) using public tools. The number holds — but Dee