DeepSeek V4-Flash exits preview, beats Pro

DeepSeek moved DeepSeek-V4-Flash-0731 out of preview and into public beta on July 31, the official release of an API that had been running under a preview tag. The model name and calling method are unchanged, deepseek-v4-flash now points at the new build, and pricing holds at $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens. The architecture and size are identical to the preview, 284 billion total parameters with 13 billion active per token and a one-million-token context window; DeepSeek says only the post-training process was repeated.

The benchmark jump is the news. DeepSeek’s own published numbers put V4-Flash-0731 ahead of its flagship V4-Pro-Preview on all nine agent and coding benchmarks the company reports, including a Terminal-Bench 2.1 score of 82.7 against 72.1 for V4-Pro-Preview and 61.8 for the earlier Flash preview. The new build also natively supports OpenAI’s Responses API format and is specifically adapted for Codex. DeepSeek says this update touches only the Flash API; V4-Pro and the models behind DeepSeek’s consumer app and website are unchanged for now, with an official V4-Pro release still to come.

What it means for operators

A cheaper tier beating the flagship tier on agent benchmarks is an unusual result worth verifying against your own workload before switching, since published benchmark gains do not always transfer to a specific agent harness or tool-calling pattern. But if it holds up, it changes the cost calculus for anyone routing coding-agent or tool-use traffic to DeepSeek’s Pro tier specifically for capability: the same money now buys strictly worse published numbers than the Flash tier offers at its existing price. Given DeepSeek’s Flash was already priced for volume use, this is one more data point in the broader price-performance war among model vendors, where capability keeps arriving at the cheap tier faster than pricing catches up.