DeepSeek-V4-Flash entered public beta on 2026-07-31 with significantly enhanced agent capabilities, outperforming V4-Pro-Preview across benchmarks: Terminal Bench 2.1: 82.7, NL2Repo: 54.2, Cybergym: 76.7, Deep SWE: 54.4. The model natively supports Responses API format and is adapted for Codex.

On 2026-04-24, V4-Pro and V4-Flash launched via OpenAI and Anthropic interfaces. Legacy models deepseek-chat and deepseek-reasoner will be discontinued by 2026-07-24, currently mapping to V4-Flash modes.

Throughout 2025, iterative upgrades included V3.2, V3.1-Terminus, and V3.1 with hybrid reasoning architecture. V3.1 achieved SWE-bench Verified: 66.0, improved reasoning efficiency, and enhanced agent capabilities. The R1-0528 reasoning model (2025-05-28) boosted AIME 2025 to 87.5 and GPQA to 81.0.

Earlier 2024 releases introduced V3 (2024-12-26), V2.5-1210 with MATH-500 at 82.8%, and the merged V2.5 (2024-09-05) achieving HumanEval 89%. The API launched context caching on disk (2024-08-02), reducing prices by an order of magnitude.