DeepSeek Releases V4 Flash With Agentic Gains

DeepSeek's model card reports sizable gains over the preview V4 Flash across its listed coding and agent benchmarks. On the published table, Flash-0731 scored 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. Those results exceed the listed V4 Pro preview scores on each of those tests.

The company notes that its code-agent evaluations used its DeepSeek Harness in a minimal configuration, max reasoning effort, temperature 1.0, and top-p 0.95. It also labels DSBench-FullStack and DSBench-Hard as internal development test sets, an important distinction for teams assessing how much weight to place on those numbers.

Independent measurements from Artificial Analysis put Flash-0731 at 50 on its Intelligence Index, versus 40 for the preceding V4 Flash and 44 for V4 Pro. Artificial Analysis reported an Elo score of 1,559 on GDPval-AA v2, its agentic real-world work-task evaluation, compared with 1,189 for the earlier Flash release. Its separate Terminal-Bench 2.1 measurement was 79%.

Artificial Analysis reported that Flash-0731 sits on its intelligence-versus-cost Pareto frontier and estimated its first-party API cost per task at about 60% below GPT-5.6 Luna (max), after accounting for OpenAI's reported price reduction. The firm attributed a major part of that result to DeepSeek's approximately 98% cache-hit discount, compared with the 90% cache-hit discount it described as common across the industry.

The same assessment found that the model's AA-Omniscience score improved by seven points relative to the previous Flash release, which Artificial Analysis attributed to a lower hallucination rate rather than higher overall accuracy. It reported an 84% hallucination rate, down 12 points from the prior model. Benchmark definitions and task mixes differ substantially, so teams comparing providers commonly need workload-specific evaluations for tool use, retrieval, long-context behavior, latency, and failure handling rather than relying on a single aggregate score.

DeepSeek's Responses API documentation currently lists support for deepseek-v4-flash only. The documentation states that deepseek-v4-pro support is due in early August 2026. The API uses the OpenAI SDK with DeepSeek's https://api.deepseek.com base URL, and supports streaming via semantic server-sent events rather than an OpenAI-style data: [DONE] terminator.

For existing integrations, the endpoint alias change makes model pinning and regression testing particularly relevant. Teams that require reproducible outputs generally evaluate named model versions, reasoning settings, cache behavior, and streaming-event handling before moving production workloads to a newly mapped endpoint.