DeepSeek-V4-Flash Official API Released, Agent Performance Nears Opus 4.8
机器之心 · wechat · 2026-07-31
The official API for DeepSeek-V4-Flash is now in public beta. Following extensive post-training, the new version significantly enhances agent capabilities and instruction following, with native support for the Responses API and Codex optimization.
V4-Flash delivers surprising results across multiple benchmarks: scoring 82.7 on TerminalBench2.1, 76.7 on Cybergym, 25.1 on AutomationBench, and 68.7 on DSBench-FullStack. Its overall agent performance is approaching that of Opus 4.8, but at just 1/90th of the price of Claude.
This upgrade applies exclusively to the V4-Flash API, leaving the App, Web, and Pro API versions unchanged for now. The official team also hinted that the more powerful DeepSeek-V4-Pro is coming soon.
More from Models
- DeepSeek V4-Flash Beats GPT-5.6 Luna in Multimodal Canvas Tests at Same Price — teortaxesTex · 2026-07-31
- DeepSeek V4-Flash Performance Jump May Stem from V4-Pro as RL Teacher — teortaxesTex · 2026-07-31
- DeepSeek Praised for World-Class RL Training That Avoids Hallucinations — teortaxesTex · 2026-07-31
- Hands-on: OpenAI's o3 Remains a Beast for OSINT Tasks — bytebot · 2026-07-31
- Rumor: Gemini 3.5 Pro Could Drop This Weekend — gaganghotra_ · 2026-07-31
- Kimi's RL Training Praised: Effectively Avoids Hallucinations and Verbosity — teortaxesTex · 2026-07-31