DeepSeek announced on 2026-08-13 that DeepSeek-V4-Pro had reached general availability, alongside its lower-cost sibling V4-Flash. The company framed the release around agent work rather than chat, describing what it called major agent upgrades with production gains, and it exposed reasoning effort as an explicit control with three settings: low for simple tasks, high for daily agent workflows, and max for complex tasks. API model names did not change, so existing integrations pick up the new behavior without code edits. On the app and web clients, V4-Pro is reached through a setting DeepSeek calls Expert Mode.
The release also adds native support for the OpenAI Responses API, which DeepSeek says is optimized for use with Codex. That matters more than it sounds: the Responses API is the interface most agent harnesses now target, and a Chinese-hosted model that speaks it natively can be dropped into an existing agent stack as a swap rather than a port.
The pricing change is the part with immediate budget consequences. New API rates took effect 2026-08-16 at 16:00 UTC and introduce a peak and off-peak split, with off-peak rates 50 percent lower than peak. DeepSeek defines peak hours as 01:00 to 04:00 and 06:00 to 10:00 UTC, with every other hour off-peak. For V4-Pro the published rates are 0.66 dollars per million input tokens off-peak and 1.32 dollars peak on a cache miss, 0.022 dollars off-peak and 0.044 dollars peak on a cache hit, and 1.98 dollars off-peak against 3.96 dollars peak for output. V4-Flash runs at 0.22 dollars off-peak and 0.44 dollars peak for input on a cache miss, 0.007 dollars and 0.014 dollars on a cache hit, and 0.66 dollars against 1.32 dollars for output.
For anyone running batch or overnight agent workloads, the peak and off-peak structure turns scheduling into a direct cost lever rather than an operational detail. Combined with cache-hit pricing that is roughly thirty times cheaper than a cache miss, the economics now reward workloads that both reuse prompt prefixes and run outside two narrow UTC windows. It is also a signal about how frontier-adjacent inference capacity is being rationed: rather than raise headline prices, DeepSeek is pricing the clock.