FULL STORY

DeepSeek-V4-Flash Launches Amid Benchmark and Pricing Debates

DeepSeek launched the V4-Flash API, with leaked benchmarks revealing top-tier performance. However, its aggressive pricing strategy quickly sparked debates within the community.

2026-07-31 ~ 2026-07-31 · 4 episodes · 34 posts

Episode 1 · DeepSeek Launches V4-Flash API with Major Agent Upgrades (2026-07-31, 21 posts)

According to official changelogs and API documentation, the stable version of DeepSeek-V4-Flash is now available via API and open for public beta. The official release of V4-Pro is also teased. The core highlight of this update is a massive leap in Agent capabilities, with benchmark scores significantly surpassing the previous V4-Pro-Preview version, making it a crucial update for developers.

已确认

  • The stable version of DeepSeek-V4-Flash is now accessible via the official API and has entered public beta.
  • Officials have teased the upcoming official release of DeepSeek-V4-Pro.
  • It is confirmed that the Agent capabilities of the stable V4-Flash version have been substantially improved.
  • In multiple benchmarks, the stable V4-Flash scores significantly higher than the previous V4-Pro-Preview, showing excellent performance in tests like TerminalBench2.1.
  • Leaked internal evaluation sets include DSBench-FullStack for assessing full-stack development capabilities, indicating significant improvements over the Preview version.

为什么重要

  • The substantial boost in Agent capabilities marks a strengthened practicality and competitiveness for DeepSeek in agentic use cases, providing developers with a more powerful model for tool calling and task execution.

1 more related posts →

Episode 2 · DeepSeek V4-Flash Scores 50 on Intelligence Index, Surpassing Pro Preview (2026-07-31, 7 posts)

Benchmark results for the official DeepSeek-V4-Flash-0731 release are in. It scored 50 on the Artificial Analysis Intelligence Index, jumping 10 points from the previous generation and outperforming its own V4-Pro-Preview by 6 points. Nearing the scores of top-tier industry models, it demonstrates exceptional cost-effectiveness and has drawn widespread community attention to DeepSeek's rapid R&D pace.

Confirmed

  • Artificial Analysis evaluation report confirms DeepSeek-V4-Flash-0731 achieved an intelligence index of 50
  • This score marks a 10-point improvement over the previous generation, beating DeepSeek-V4-Pro-Preview by 6 points
  • It trails only 1 point behind GLM-5.2 and GPT-5.6 Luna
  • It performs exceptionally well on the intelligence-cost Pareto frontier, offering better cost-effectiveness than the newly discounted (80% off) GPT-5.6 Luna (xhigh)

Why it matters

  • In just a month and a half of iterative updates, DeepSeek's capabilities have nearly matched top-tier models like GLM 5.2, showcasing惊人的 R&D speed
  • The Flash version's benchmark score significantly surpassing the Pro preview indicates a major breakthrough in DeepSeek's technical approach
  • Even with GPT-5.6 Luna announcing an 80% price cut, DeepSeek-V4-Flash still wins on cost-effectiveness, proving highly competitive
  • It continues to maintain its position in the core tier of the AI race, showing that its low-profile yet rapid iteration strategy is highly effective

Episode 3 · DeepSeek-V4-Flash Benchmarks Impress but Face Pricing Controversy (2026-07-31, 4 posts)

DeepSeek-V4-Flash shows impressive benchmark scores and low inference costs, but developers spotted abnormally high cache write fees in its API pricing. If this billing anomaly is removed, the model's cost-effectiveness significantly outperforms competitors.

Episode 4 · Predictions and Buzz Around DeepSeek V4 (2026-07-31, 2 posts)

Amidst a fierce AI price war, predictions are swirling around DeepSeek's next-generation V4-PRO model, which is expected to rival Claude Opus 4.8 in performance while continuing to break industry pricing floors.