DeepSeek V4 Flash, launched on July 31, topped multiple AI leaderboards while being priced at roughly one-tenth the cost of comparable models at $0.14 per million input tokens. Independent testing by Composio on 30 complex multi-step workflows using real tools like Gmail, GitHub, Slack, and Google Sheets showed the model completed only 53.8% of tasks, with 129 successes out of 240 runs. Among eight agent harnesses tested, Pi Agent performed best with 20 out of 30 tasks completed, highlighting how integration significantly impacts results. The current beta label and scheduled pricing adjustment on August 16, 2026 suggest the model is not yet ready for mission-critical deployment.
Source: Read the original article
