DeepSeek V4-Flash Official Release: Benchmarks Near Opus 4.8 at 1/90 Claude Price

DeepSeek announced that the V4‑Flash model has been updated to an official version and is now available for public testing. According to the official documentation, the Flash release shows markedly higher test results than the previous V4‑Pro‑Preview version.

The new model benefits from extensive post‑training, which greatly enhances its agent capabilities and instruction‑following performance, making it far more practical for development tasks.

Benchmark results include:

Terminal Bench 2.1 (terminal operation) – score 82.7

NL2Repo (code repository understanding) – score 54.2

DeepSWE (code modification) – score 54.4

Cybergym (cyber‑security tasks) – score 76.7

Toolathlon Verified (tool‑calling) – score 70.3

In broader agent evaluations, DeepSeek‑V4‑Flash achieved 25.2 on Agent Last Exam and 25.1 on Automation Bench Public. On full‑stack development benchmarks, it scored 68.7 on DSBench‑FullStack and 59.6 on the harder DSBench‑Hard.

From these results, the Flash version’s performance on multiple agent benchmarks is approaching that of Opus 4.8. Notably, the Flash model is only 284 B in size and is priced at roughly one‑ninetieth of Claude, dramatically lowering the entry barrier for large‑model usage.

The release also adds native support for the Responses API and includes optimizations for Codex. The model architecture and size remain identical to the V4‑Flash‑preview; the improvement stems solely from additional fine‑tuning.

It is important to note that this upgrade applies only to the DeepSeek‑V4‑Flash API; the V4‑Pro API and the App/Web models remain unchanged. The community has reacted strongly, noting the dramatic performance jump and speculating that DeepSeek’s pricing strategy may be a response to recent OpenAI price reductions.

Looking ahead, DeepSeek hinted that the V4‑Pro version is forthcoming, prompting excitement about the potential capabilities of the next iteration.