On Aug 13, 2026, Chinese AI company DeepSeek officially released the general-availability version of DeepSeek-V4-Pro (codename V4-Pro-0813). The model focuses on strengthening Agent capabilities, supports adjustable inference intensity, and is natively compatible with the OpenAI Responses API. At the same time, DeepSeek announced it will apply busy/idle segmented pricing (peak-off-peak pricing) to the V4-series API; off-peak rates will be half of peak rates. The new prices take effect on Aug 16 at 16:00 UTC (Beijing Time Aug 17 00:00).

Model details

The V4-Pro production release builds on the preview with post-training improvements that significantly raise production-level Agent performance. Key disclosed metrics include:

Terminal Bench 2.1: score 87.9
DeepSWE: score 62.7
Other Agent-related benchmarks (such as CyberGym, NL2Repo) also show substantial gains

The model supports a 1M token context window and a maximum output length of 384K tokens, offering switchable reasoning (thinking) and non-reasoning modes. Users can experience this via an "Expert Mode" on the App/Web; the API model name remains unchanged.

New feature highlights:

Adjustable inference intensity: three tiers—low (simple tasks), high (daily Agent workflows), and max (complex tasks)—available for V4-Pro and V4-Flash.
Native support for the OpenAI Responses API, optimized for tools like Codex, enabling one-click configuration.
Support for JSON output, tool calling, and Anthropic API compatibility.

The V4-Flash production release was launched at the end of July and performed well in some evaluations. The V4-Pro GA release addresses higher-end capability gaps, forming a dual-track strategy with "Pro for high-intelligence Agents, Flash for cost-effective efficiency."

Details of busy/idle segmented pricing

DeepSeek simultaneously adjusted API pricing, introducing a peak-off-peak mechanism to optimize resource allocation and service stability:

Peak periods (Beijing Time 09:00–12:00, 14:00–18:00; corresponding to UTC 01:00–04:00, 06:00–10:00):

V4-Pro: cache-hit input 0.30 CNY per million tokens, cache-miss input 9.0 CNY, output 27.0 CNY
V4-Flash: cache-hit 0.10 CNY, cache-miss 3.0 CNY, output 9.0 CNY

Off-peak periods (all other times): prices are halved.

The new pricing is generally higher than previous levels, but the peak-off-peak differential is intended to encourage users to schedule non-urgent tasks during low-demand periods, lowering average costs and alleviating peak compute pressure. Cache hits still receive substantial discounts.

Potential impacts: cost structure and market competition

1. Impact on developers and enterprises The substantial Agent capability improvements make V4-Pro more suitable for complex multi-step workflows, code generation and tool-invocation scenarios. Busy/idle segmented pricing requires users to optimize call strategies: high-priority tasks may accept peak premiums, while batch or delay-tolerant tasks shift to off-peak windows to control costs. Compatibility with the OpenAI ecosystem lowers migration barriers.

2. Impact on DeepSeek The GA release helps consolidate its position in cost-effective and open-weight model segments, while the pricing changes aim to improve revenue structure to support compute expansion and continued R&D. The earlier phenomenon of V4-Flash outperforming the Pro preview in some tests also prompted accelerated refinement of Pro.

3. Impact on industry competition V4-Pro's strong Agent performance intensifies competition with domestic models such as Kimi K3 and Qwen, while offering near-leading capabilities at prices far below OpenAI an