DeepSeek V4 Flash Is Official: Major Agent Gains and Native Codex Support
DeepSeek-V4-Flash-0731 beats V4-Pro-Preview across nine vendor-reported agent benchmarks, adds native Responses API support for Codex, and keeps the same model architecture.
DeepSeek released the official DeepSeek-V4-Flash API into public beta on July 31, 2026. The call pattern stays the same: developers select deepseek-v4-flash and receive the updated 0731 model.
DeepSeek now presents Flash as an agent and coding model for serious tool-driven work. Speed and low cost remain part of the offer, but they are no longer the whole story.
Agent scores move beyond V4-Pro-Preview
According to DeepSeek's published results, the official Flash build beats V4-Pro-Preview on all nine listed agent and coding agent benchmarks. It scores 82.7 on Terminal Bench 2.1, compared with 72.1 for V4-Pro-Preview and 61.8 for Flash Preview. On DeepSWE, the new score is 54.4, up from 7.3 for Flash Preview. DSBench-FullStack rises from 37.0 to 68.7.
The same table puts V4 Flash above GLM-5.2 on every row where both models have a score. Several results sit close to Opus-4.8, including Terminal Bench 2.1 at 82.7 versus 85.0 and Agents' Last Exam at 25.2 versus 25.7.
These are DeepSeek's numbers, not independent results. For public Code Agent tasks, DeepSeek used the upcoming DeepSeek Harness in minimal mode with max effort, top_p=0.95, and temperature=1.0. DSBench-FullStack and DSBench-Hard are internal test sets. Different harnesses and run settings make the cross-provider comparisons useful as a direction, not a controlled ranking.
Native Responses API support brings Flash to Codex
For developers, protocol support is the clearest practical change. V4 Flash now supports the Responses API natively and has been adapted specifically for Codex.
DeepSeek's setup guide says one configuration can expose the model across Codex CLI, the ChatGPT desktop app, and the Codex extension for VS Code. At release, Flash is the only DeepSeek model with documented Codex support.
That integration gives developers a first-party route for using DeepSeek inside a coding agent that can work with files, terminal commands, tools, and longer task loops. It also removes the need to force Codex through a Chat Completions compatibility layer.
The model size did not change
DeepSeek-V4-Flash-0731 uses the same architecture and model size as DeepSeek-V4-Flash-Preview. DeepSeek says it only repeated the post-training process.
The release therefore attributes the large agent gains to post-training and agent adaptation, not a larger checkpoint. That detail matters for teams evaluating whether Flash can deliver stronger coding performance while keeping its existing efficiency profile.
What changed, and what did not
This is an API-only update. DeepSeek-V4-Pro API and the models used in DeepSeek's App and Web experience have not changed. DeepSeek says the official V4 Pro release will follow soon.
The narrow release boundary is important. The July 31 announcement does not make every V4 surface a new production build, and it does not replace the separate Pro roadmap.
Flash now has a clearer job in the V4 family. It is DeepSeek's first production V4 model with native Codex support and a benchmark package aimed directly at coding agents.