XNews

DeepSeek V4-Flash-0731 Official API Release with Enhanced Agent Capabilities

DeepSeek V4-Flash-0731 Official API Release with Enhanced Agent Capabilities

DeepSeek upgraded V4-Flash from preview to official API version 0731 on July 31, 2026. The model architecture stayed at 284B total parameters with 13B activated via mixture-of-experts; all changes came from post-training. Key additions: Agent capabilities now surpass the previously stronger V4-Pro-Preview on multiple benchmarks, including Terminal Bench 2.1 at 82.7 and DeepSWE at 54.4-54.5 (close to Claude Opus 4.8). It also beats GLM-5.2 and is near GPT-5.6 Luna. The API now natively supports the Responses API format, specifically adapted for OpenAI Codex, removing the need for proxy conversions and offering a one-click configuration script. Pricing remains $0.14/1M input tokens, $0.28/1M output, with a 1M-token context window—an order of magnitude cheaper than comparable OpenAI models. DeepSeek Harness is upcoming and internal benchmarks DSBench-FullStack and DSBench-Hard were teased. The App and web versions are unchanged; V4-Pro official release date unknown. Third-party platform Cola added Flash to its Token Plan.

Key insights

Post-training alone lifted the smaller Flash model above the larger Pro preview in agent tasks

V4-Flash-0731 retains the same 284B total/13B active MoE architecture as before. After post-training, its agent benchmarks—such as Terminal Bench 2.1 (82.7) and DeepSWE (54.4-54.5)—exceed those of V4-Pro-Preview, and come close to Claude Opus 4.8. This shows that for agent-oriented tasks, improved training techniques can outpace model scale, without adding parameters.

V4-Flash now natively supports OpenAI Codex through Responses API format

DeepSeek added Responses API format compatibility in V4-Flash-0731. This format is tailored for OpenAI Codex, which runs across CLI, ChatGPT desktop, and VS Code. Previously, developers needed intermediary proxies; now a one-click config script switches all Codex clients to DeepSeek. This integration expands DeepSeek’s reach into the AI‑coding assistant ecosystem.

Flash offers a drastically lower cost structure for agent‑heavy coding tasks

At $0.14 per million input tokens and $0.28 per million output, with a 1M‑token context window, DeepSeek V4-Flash is an order of magnitude cheaper than comparable OpenAI models used by Codex. For coding loops that consume large token volumes, this pricing changes the economic calculus, making large‑scale autonomous agent runs more feasible.

Flash surpasses GLM‑5.2 and approaches GPT‑5.6 Luna in overall intelligence

According to both DeepSeek’s own benchmarks and third‑party evaluations, V4‑Flash matches or exceeds GLM‑5.2 and is near GPT‑5.6 Luna. The model has been listed on the Cola Token Plan, offering free access for testing, which indicates early community validation.

Source posts