DeepSeek V4-Flash Official API Enters Public Beta: A Post-Training Leap in Agent Capabilities

On July 31, 2026, DeepSeek announced via its API docs changelog that the official DeepSeek-V4-Flash API (codenamed V4-Flash-0731) had entered public beta. Model structure and parameter counts are identical to the preview; a fresh post-training and alignment pass delivered a major jump in agent capabilities — Terminal Bench 2.1 at 82.7 and DeepSWE jumping from 7.3 to 54.4.

Same Architecture, Gains From Post-Training Alone#

• MoE architecture with 284B total / 13B active parameters and native 1M-token context, identical to the V4-Flash-preview released April 24
• The upgrade is purely a new post-training and alignment pass — no larger parameter count
• All 9 official Code Agent benchmarks beat the V4-Pro-Preview: Terminal Bench 2.1 82.7, Cybergym 76.7, Toolathlon verified 70.3, DSBench-FullStack 68.7, DSBench-Hard 59.6, DeepSWE 7.3 → 54.4
• Evaluated with DeepSeek Harness minimal mode, max effort, top_p=0.95, temperature=1.0; overall it outpaces V4-Pro-Preview and GLM-5.2, approaching Claude Opus 4.8 on some metrics

Native Responses API, Tuned for Codex#

• The official version natively supports the Responses API format and is specifically adapted for the Codex ecosystem
• One configuration lets developers call V4-Flash directly from Codex CLI, the ChatGPT desktop app, and the VS Code Codex plugin
• Existing Chat Completions / Anthropic-compatible endpoints keep working; Base URL and model name (deepseek-v4-flash) are unchanged, so current deployments can switch without edits

Flash-Only Update; Pro Stable Is Next#

• This update covers only the V4-Flash API; the V4-Pro API and App/Web models are unchanged
• The legacy model names deepseek-chat / deepseek-reasoner were retired on July 24 and now map to V4-Flash's non-thinking and thinking modes
• DeepSeek said at the time that the V4-Pro official release would follow soon. As of this article's September 21, 2026 update that has since happened and moved on: DeepSeek's own API changelog records V4-Pro shipping on August 13 (0813), and V4.1 Flash arriving on September 10 (see this site's V4.1 Flash piece, which is filed as analysis because no first-hand DeepSeek announcement could be verified)

Our Take#

• The headline is "no parameter bloat": the 284B/13B scale is unchanged and the capability gains come entirely from post-training quality
• The shift from generic chat to agent infrastructure — Responses API plus Codex adaptation — targets real development workflows rather than single-turn Q&A
• Pricing is unchanged (0.02 CNY/M tokens on cache hits, 1 CNY/M on misses, 2 CNY/M output), a genuine win for developers; DeepSeek has previewed peak/off-peak pricing that would double prices during rush hours, effective date TBD