DeepSeek-V4-Flash reshapes the Arena cost-performance frontier
August 5, 2026Posted by Federico Ulfo

Agent Arena: DeepSeek-V4-Flash (High) lands on the cost-performance Pareto frontier with a $0.024 median cost per task, to the right of GPT-5.6 Luna (xHigh) at $0.026.
Frontend Code Arena: a 1586 score at $0.14/$0.28 per MToken — the best performance-per-dollar of any model on the board.
Context: DeepSeek's V4-Flash API public beta was already covered in an earlier update ("DeepSeek opens the V4 Flash API, and its agent scores jump"); these Arena results extend that story.
Sources on X: