DeepSeek V4-Flash-0731 in public beta: the Codex protocol arrives at the Chinese rival

In plain terms

DeepSeek releases V4-Flash-0731 in public beta on July 31, with an explicit "agent tasks" framework - Responses API, Codex adaptation, two announced scores (82.7 on Terminal Bench 2.1, 54.4 on DeepSWE). In other words: the Chinese lab exits the "general model" race and delivers a competitive variant in the field that matters today, the looped agent.

What really changes

V4-Flash-0731, documented on api-docs.deepseek.com/updates (July 31, 2026), is not an architectural overhaul but a post-training patch focused on three things: tool-use robustness, compliance with the Responses format (the OpenAI protocol adopted by Codex), and long-running execution with state. The displayed scores place V4-Flash on a useful plateau, one notch below the closed borders as of their respective communication dates.

Under the hood

Terminal Bench 2.1 measures the ability to execute shell/dev tasks in a controlled environment over dozens of turns; DeepSWE measures the resolution of real GitHub issues. A model that holds 82 on Terminal Bench without derailing is usable as a Codex runtime without a massive wrapper. The decisive technical point: V4-Flash implements the Responses format, so it plugs into an existing Codex chain by changing the provider's URL - that's the real "so what."

So what

Two implications. Codex / Claude Code practitioners: for the first time, a Chinese provider offers credible protocol compatibility - the runtime vendor lock decreases, price/latency arbitration becomes possible without rewriting the orchestration. Open-model-economics thread: this is the direct counterexample to the "6 months to live" thesis - DeepSeek continues to release aggressive pricing variants on the same tracks as closed borders, without announcing a pivot. To watch: real latency of V4-Flash agents in prolonged tool-use, behavior in case of rate-limiting on DeepSeek's side, and propagation to orchestrators (LangChain, LlamaIndex, Cursor, Continue).