DeepSeek has put V4 Flash into public beta, according to a post the company quoted as announcing that the "official API is now LIVE" and that agent capabilities were "massively upgraded." DeepSeek also claims the model now natively supports the Responses API format and is "fully adapted for Codex," with configuration details linked in its official API docs.
The accompanying comparison places DeepSeek-V4-Flash-0731 against V4-Flash-Preview, V4-Pro-Preview, GLM-5.2 and Opus-4.8 across a set of code and agent benchmarks. In that table, the new version posts higher scores than DeepSeek-V4-Flash-Preview and DeepSeek-V4-Pro-Preview on every listed measure, including Terminal Bench 2.1 at 82.7, NL2Repo at 54.2, DeepSWE at 54.4 and DSBench-Hard at 59.6. It also trails Opus-4.8 in several rows, including Terminal Bench 2.1, NL2Repo, Cybergym, DeepSWE, Toolathlon-Verified, Agents’ Last Exam, AutomationBench (Public), DSBench-FullStack and DSBench-Hard.
Fine print below the table states that public Code Agent tasks were tested with an upcoming DeepSeek Harness "minimal mode" framework using max tier, topp=0.95 and temperature=1.0. DeepSeek also describes DSBench-FullStack and DSBench-Hard as internal benchmark sets, so the numbers should be treated as company-reported results rather than independently verified comparisons.
Source: DeepSeek AI
