Updated
Updated · Forbes · Aug 3
DeepSeek Launches V4 Flash at $0.03 per Task as 37% Accuracy Clouds 105x Cost Edge
Updated
Updated · Forbes · Aug 3

DeepSeek Launches V4 Flash at $0.03 per Task as 37% Accuracy Clouds 105x Cost Edge

3 articles · Updated · Forbes · Aug 3

Summary

  • DeepSeek’s V4 Flash debuted Friday at an estimated $0.03 per benchmark task, versus $3.15 for Anthropic’s Claude Fable 5, giving it a roughly 105-fold price advantage.
  • Artificial Analysis scored V4 Flash at 50 on its Intelligence Index—matching Google Gemini 3.6 Flash—but its AA-Omniscience benchmark showed just 37% accuracy and an 84% hallucination rate.
  • Those results suggest the savings could be offset by retries, human verification, compliance checks, delayed projects and errors that reach customers or employees.
  • OpenAI’s July 17 “Useful Intelligence Per Dollar” framework offers a way to judge that trade-off by measuring successful tasks, true cost per task, accuracy and whether quality holds as usage scales.
  • For employers, the broader question is whether cheaper models fit only low-risk workflows, while higher-risk work still justifies paying more for stronger output quality.

Insights

Could the world's cheapest AI model actually bankrupt your business through hidden labor and compliance costs?
If a budget AI hallucinates most of the time, are employees becoming mere babysitters for flawed algorithms?
Why are tech giants secretly terrified of a free, open-weight model despite its massive accuracy flaws?