DeepSeek-V4-Flash exits preview, ships official public beta with major agent-benchmark gains

DeepSeek moved DeepSeek-V4-Flash out of preview and into an official public beta on July 31, re-post-training the same architecture to deliver what the company describes as significantly enhanced agent capabilities over the prior preview build.

What's new

DeepSeek's API changelog states plainly: "The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling method remains unchanged — simply set the model name to deepseek-v4-flash to use the latest version." The new build, internally referred to as DeepSeek-V4-Flash-0731, keeps the same model architecture and parameter size as DeepSeek-V4-Flash-Preview — this is a re-post-training pass, not a new base model.

DeepSeek published a wide benchmark sweep showing gains it says "far exceed V4-Pro-Preview" on agentic and coding tasks specifically:

• Terminal Bench 2.1: 82.7
• NL2Repo: 54.2
• Cybergym: 76.7
• DeepSWE: 54.4
• Toolathlon (verified): 70.3
• Agent Last Exam: 25.2
• Automation Bench (Public): 25.1
• DSBench-FullStack: 68.7
• DSBench-Hard: 59.6

The company notes the code-agent benchmark figures were measured using an as-yet-unreleased "DeepSeek Harness" in minimal mode, at max effort level with top-p 0.95 and temperature 1.0 — a detail worth flagging for anyone trying to reproduce the numbers before that harness ships. The official V4-Flash release also adds native support for the Responses API format, with configuration specifically adapted for use with Codex. DeepSeek was explicit that this update is scoped to the API only — the DeepSeek-V4-Pro API and the consumer app/web models are unchanged — and said an official V4-Pro release will follow "soon."

Context

DeepSeek shipped V4-Pro and V4-Flash together as preview models in April 2026, offered through both an OpenAI-compatible ChatCompletions interface and an Anthropic-compatible interface, with the older deepseek-chat/deepseek-reasoner model names retired as of July 24. This release is the first of the two April preview models to graduate to an official, numbered build, with V4-Pro still pending. The emphasis on agent and tool-use benchmarks specifically — rather than general knowledge or reasoning scores — continues a pattern set by V3.1 and V3.2, where DeepSeek has repeatedly targeted coding-agent and tool-calling performance as its main point of competitive differentiation against frontier-priced US labs.

Why it matters

DeepSeek's models compete primarily on cost-per-capability, and this release sharpens that pitch specifically for the agentic coding workloads increasingly driving enterprise API spend — Terminal Bench, SWE-style, and tool-orchestration benchmarks are the numbers developers now check first when picking a model for autonomous coding agents, not general knowledge leaderboards. Native Responses API support and explicit Codex-oriented configuration also signal DeepSeek is optimizing for drop-in compatibility with the tooling ecosystem OpenAI has built up around agentic coding, lowering the switching cost for teams already standardized on that interface. With V4-Pro still to come, this is a partial rollout — the more capability-focused half of the V4 family remains on preview footing for now.