Release Overview
<cite index="20-1,20-2">The official release of the DeepSeek-V4-Flash application programming interface (API) entered public beta on July 31, 2026, with the calling method remaining unchanged — developers set the model name to `deepseek-v4-flash` to access the latest version.</cite> The build is designated DeepSeek-V4-Flash-0731.
<cite index="17-9,17-10,17-11">The release is a retrained build, not a new architecture: DeepSeek-V4-Flash-0731 keeps the same model architecture and size as the preview, with the upgrade consisting of re-post-training — new weights on the same skeleton.</cite>
Architecture and Specifications
<cite index="2-5,2-6,2-7">DeepSeek V4 Flash is an efficiency-optimised Mixture-of-Experts (MoE) Large Language Model (LLM) with 284 billion total parameters and 13 billion activated parameters, supporting a one-million-token context window, designed for fast inference and high-throughput workloads while maintaining strong reasoning and coding performance. The model includes hybrid attention for efficient long-context processing.</cite> <cite index="2-2">Maximum output is 384,000 tokens.</cite>
Benchmark Results
<cite index="20-5">According to DeepSeek's official API changelog, the 0731 build posts the following vendor-stated scores: Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, Cybergym at 76.7, DeepSWE at 54.4, Toolathlon (verified) at 70.3, Agent Last Exam at 25.2, Automation Bench (Public) at 25.1, DSBench-FullStack at 68.7, and DSBench-Hard at 59.6.</cite>
<cite index="3-5">The 0731 build scores higher than DeepSeek's own V4-Pro-Preview on all nine agent and coding benchmarks the company published.</cite> On Terminal Bench 2.1, <cite index="6-8,6-9">V4-Flash-0731 scored 82.7 compared to V4-Pro-Preview's 72.1, a 14.7% win for the budget model over its premium sibling.</cite> <cite index="21-1">On DeepSWE and DSBench-FullStack, the 0731 build scored 54.4 and 68.7 respectively, compared to 7.3 and 37.0 for the preview.</cite>
An important methodological caveat applies: <cite index="1-10">these are vendor-published numbers, produced with the vendor's own harness at max effort settings.</cite> <cite index="19-12,19-13">Two of the listed benchmarks — DSBench-FullStack and DSBench-Hard — are internal test sets, and independent numbers were expected to take several days to appear.</cite> Separately, <cite index="5-4">Artificial Analysis, an independent evaluator, scores DeepSeek V4 Flash 0731 at 50 on its Intelligence Index, placing it well above average among open-weight models of similar size, against a median of 25.</cite>
Pricing
<cite index="4-1,4-2">Official pricing per one million tokens stands at $0.0028 for cache-hit input, $0.14 for cache-miss input, and $0.28 for output. V4-Pro's unchanged preview rates are $0.003625 / $0.435 / $0.87.</cite> <cite index="5-5">Artificial Analysis independently confirms $0.14 per one million input tokens — rated very competitive against a sector median of $0.58 — and $0.28 per one million output tokens against a median of $2.20.</cite>
Developers should note one cost nuance: <cite index="9-1,9-2">thinking mode is on by default in V4-Flash, and thinking tokens — the internal reasoning steps the model generates — are billed at the output token rate even though they do not appear in the final response, which can materially increase effective cost per call above the headline $0.14 per million input tokens for agent tasks that trigger extended reasoning chains.</cite>
New Capabilities and Ecosystem Integration
<cite index="23-4">The post-training upgrade significantly improves agentic, coding, and tool-calling ability and adds native Responses API support with Codex adaptation.</cite> <cite index="22-11,22-12">OpenAI's Codex talks to models through the Responses API, which 0731 now supports natively, covering the Codex command-line interface, desktop app, and integrated development environment extension; as of the release, only Flash supports Codex, with V4-Pro support expected in early August 2026.</cite>
Context and Outlook
<cite index="4-10">The release tracks a broader 2026 pattern where labs increasingly treat agentic workloads — high-volume, tool-calling, parallel — as the product, routing them to smaller models where post-training quality matters more than parameter count.</cite> <cite index="9-7">DeepSeek signalled in the same announcement that the official release of V4-Pro will also follow "as soon as possible."</cite> <cite index="22-5">The MIT-licensed weights on Hugging Face remain at the April preview, meaning the version available for local download is the older checkpoint as of publication.</cite>