DeepSeek-V4-Flash Goes Stable with Official Codex Support

What happened

On July 31, DeepSeek updated its official changelog: the DeepSeek-V4-Flash stable API is now in public beta. The preview version only landed a few weeks ago, and this promotion to stable comes with one clear theme — agent capability.

Unlike the usual launch playbook that leads with parameter counts, V4-Flash stable (build 0731) shares exactly the same architecture and size as the preview. DeepSeek says the gains come almost entirely from a fresh round of post-training. The model didn't get bigger; it got better at doing actual work.

The numbers

DeepSeek published a set of coding-agent benchmark results, and the stable build beats the earlier V4-Pro-Preview across the board:

Benchmark	V4-Flash stable score
Terminal Bench 2.1	82.7
Cybergym	76.7
Toolathlon Verified	70.3
DSBench-FullStack (internal full-stack test)	68.7
DSBench-Hard (internal hard agent test)	59.6

Worth pausing on: Flash is the lightweight tier in DeepSeek's lineup. A light model outscoring the preview of the company's own flagship is the genuinely interesting part of this report card.

Official Codex support: configure once, use everywhere

The other big piece of this release: V4-Flash now natively supports the Responses API format, and DeepSeek shipped an official Codex integration. Set it up once and you can call DeepSeek models from the Codex CLI, the ChatGPT desktop app, and the Codex extension for VS Code.

There is even a one-command setup script:

macOS / Linux: bash <(curl -fsSL https://cdn.deepseek.com/api-docs/codex-deepseek-setup.sh)

Windows (PowerShell): irm https://cdn.deepseek.com/api-docs/codex-deepseek-setup-en.ps1 | iex

One prerequisite: Codex CLI or the ChatGPT desktop app must already be installed and launched at least once, so that the ~/.codex directory exists.

For now only V4-Flash can plug into Codex. V4-Pro is expected to follow in early August, and a stable V4-Pro release is also on the way.

Regular users won't notice yet

One caveat: this upgrade is API-only. It covers the V4-Flash API and nothing else — the DeepSeek web app, mobile app, and the V4-Pro API are all unchanged. If you only chat with DeepSeek in the browser, nothing feels different today.

The same day in AI, for context

Zoom out and the timing is what makes this release interesting. Within roughly the same news cycle:

OpenAI announced steep price cuts for the GPT-5.6 family: Luna's input price dropped 80% (from $1 to $0.20 per million tokens), Terra got 20% cheaper, and Sol gained a Fast mode

Zhipu opened its GLM Coding Plan to everyone — no more waiting for slots — though prices went up, with the Lite tier moving from 49 to 118 yuan per month

DeepSeek answered with the V4-Flash stable release and Codex support

Frontier models are no longer updating monthly; it feels daily now. Companies competing on models, agents, ecosystems, and price is good news for developers — the fiercer it gets, the further API prices are likely to fall.

References

DeepSeek-V4-Flash release notes: https://api-docs.deepseek.com/zh-cn/updates/

Codex setup guide: https://api-docs.deepseek.com/zh-cn/quick_start/agent_integrations/codex/