AIBRIEF
The AI benchmark firm Artificial Analysis ranks the new Flash model ahead of a 428 billion parameter rival on its intelligence vs cost curve, but independent AI developer Simon Willison's hands on test shows the model's default reasoning mode still
DeepSeek's V4-Flash-0731 model hit the top of Artificial Analysis's Intelligence Index vs. Cost per Intelligence curve on 31 July 2026, at $0.14 per million input tokens and $0.27 per million output tokens. The 304-billion-parameter model outranks a 428-billion-parameter rival on the benchmark provider's value-curve chart, putting parameter count on the wrong axis for routine-traffic routing.
For builders, the signal is direct: V4-Flash becomes the default candidate for high-volume, low-stakes work, while frontier models stay reserved for tasks that need maximum reasoning effort. The model is live on OpenRouter with adjustable reasoning_effort, and the companion paper frames the release as a push toward "million-token context intelligence" with a focus on agentic tasks.
The caveat is in the settings. Simon Willison's hands-on test drew a disappointing pelican SVG at the default reasoning level and a much better one at high reasoning, a reminder that the cheap headline still carries a configuration tax. MarkTechPost corroborates the agentic-gains framing.