DeepSeek updates V4-Flash API for coding agents without changing architecture

DeepSeek released V4-Flash-0731, an updated API for coding agents, into public beta on July 31, keeping the model's 284-billion-parameter mixture-of-experts architecture unchanged while adding native support for OpenAI's Responses API and a Codex integration guide. The update targets agentic software development, with benchmark scores including 82.7 on Terminal Bench 2.1 and 54.2 on NL2Repo, though DeepSeek noted the tests used an unreleased harness and internal benchmarks. The V4-Pro API remains unchanged, with support expected in early August 2026.

DeepSeek, the AI lab founded by Liang Wenfeng https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas?ref=runtimewire , put its updated V4-Flash API into public beta on July 31. The release uses post-training and developer tooling to reposition the smaller model for coding agents while keeping its underlying architecture and parameter count fixed, according to DeepSeek's changelog https://api-docs.deepseek.com/updates/?ref=runtimewire . DeepSeek emphasized the narrow scope of the update: V4-Flash-0731 applies only to the deepseek-v4-flash API. The V4-Pro API and the models served through DeepSeek's app and website remain unchanged. DeepSeek said the official V4-Pro release will follow, without setting a date. DeepSeek on X https://x.com/deepseek ai/status/2083084419515220191?ref=runtimewire That distinction matters because DeepSeek first released the V4 preview https://api-docs.deepseek.com/news/news260424/?ref=runtimewire on April 24th with both Flash and Pro variants. The July 31st release is a targeted update to the existing Flash line. It does not introduce a new V4 architecture or replace models across DeepSeek's product surfaces. Liang built DeepSeek out of the technical base of High-Flyer, the quantitative hedge fund he co-founded after studying engineering at Zhejiang University. His path from automated trading into frontier models helps explain the emphasis on efficiency: DeepSeek has repeatedly treated compute, inference cost and architecture as constraints to be engineered around instead of costs that disappear with scale. Same model size, different behavior DeepSeek says V4-Flash-0731 retains the preview model's mixture-of-experts architecture, with 284 billion total parameters and 13 billion activated for each token. The V4 model materials https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash?ref=runtimewire list a one-million-token context window. DeepSeek described the July build as "re-post-trained," meaning the work focused on shaping the existing model's behavior after pretraining instead of expanding the model itself. The practical target is agentic software development. V4-Flash now natively supports the Responses API, the interface used by OpenAI's Codex clients, and DeepSeek has published a dedicated Codex integration guide https://api-docs.deepseek.com/quick start/agent integrations/codex/?ref=runtimewire . Developers can configure DeepSeek as a model provider for Codex-style workflows through DeepSeek's guide. DeepSeek's documentation currently limits that integration to V4-Flash. V4-Pro support is expected in early August 2026, according to the API pricing and feature page https://api-docs.deepseek.com/quick start/pricing/?ref=runtimewire . That sequencing gives Flash a specific job: become the readily available, lower-cost DeepSeek model inside coding-agent loops while DeepSeek finishes the Pro release. DeepSeek is selling post-training as the upgrade DeepSeek reports substantial agent benchmark gains from the retuning. Its July 31st change log https://api-docs.deepseek.com/updates/?ref=runtimewire lists scores of 82.7 on Terminal Bench 2.1, 54.2 on NL2Repo, 76.7 on Cybergym, 54.4 on DeepSWE and 70.3 on Toolathlon Verified. DeepSeek says those results exceed the V4-Pro preview on several agent tasks. The comparison comes with material qualifications. DeepSeek says its code-agent tests used a "minimal mode" of an unreleased DeepSeek Harness with maximum reasoning effort, a temperature of 1.0 and top-p of 0.95. Two additional evaluations, DSBench-FullStack and DSBench-Hard, are internal DeepSeek test sets. The results show what DeepSeek optimized for, though independent testing will determine how reliably those gains transfer to real repositories and longer-running agent sessions. This release also reflects where model competition is moving. Architecture launches still draw attention, but developers experience models through tool calls, retry behavior, patch quality, context management and the cost of keeping an agent running. Post-training can improve those behaviors without funding another full pretraining run or forcing customers to migrate to a new endpoint. DeepSeek made that migration deliberately small. Existing API users keep the same base URL and select deepseek-v4-flash as the model. The compatibility choice lowers the work required to test the update inside an existing application, which can matter as much as a benchmark gain for teams already maintaining several model providers. The price advantage compounds inside agents DeepSeek currently lists V4-Flash at $0.14 per million cache-miss input tokens and $0.28 per million output tokens. Cache-hit input costs $0.0028 per million tokens. V4-Pro is priced at $0.435 for cache-miss input and $0.87 for output, with a $0.003625 cache-hit rate. Both models support one-million-token contexts and outputs of up to 384,000 tokens, according to DeepSeek's pricing documentation https://api-docs.deepseek.com/quick start/pricing/?ref=runtimewire . Those rates make Flash the clearer distribution vehicle for agent workloads. A coding agent can turn one instruction into repeated model calls for planning, file inspection, tool use, code generation and review. RuntimeWire reported on July 15th /article/deepseek-v4-pro-price-cut-agent-margin-problem that this call multiplication can erode the apparent savings from a low per-token price. Flash gives developers considerably more room before those loops become expensive. Liang's research-first strategy meets developer distribution Liang has consistently described DeepSeek as a research organization pursuing original model work. In a translated interview with Waves, he said DeepSeek wanted China to become an AI contributor instead of continuing to follow work developed elsewhere. He also described cloud services as secondary to DeepSeek's pursuit of artificial general intelligence. The V4-Flash update shows why the API still matters to that research-first strategy. Codex compatibility puts DeepSeek inside an established developer workflow, where model behavior is tested against live codebases and repeated tool interactions. Low pricing reduces the cost of experimentation. The unchanged architecture lets DeepSeek argue that the improvement came from training technique and product integration, two areas Liang can iterate on faster than a full model generation. V4-Pro remains the larger model, with 1.6 trillion total parameters and 49 billion activated parameters. Flash carries the immediate distribution work. Liang's bet is that a smaller model, trained and packaged for the way agents actually operate, can earn developer usage before DeepSeek's flagship release is ready.