DeepSeek V4-Pro API prices rise as the company releases Harness

DeepSeek V4-Pro API prices were scheduled to increase on August 16 at 4 p.m. UTC, alongside an updated V4-Pro build and the MIT-licensed developer preview of its agent software, DeepSeek Harness. For developers, the immediate change is a higher bill for V4-Pro, particularly where applications repeatedly send cached context or reread files.

The company’s V4-Pro endpoint now serves build V4-Pro-0813, according to The Decoder. DeepSeek retained the model’s name, parameter count and one-million-token context window, and said existing integrations would continue to work without modifications. The release adds native support for OpenAI’s Responses API with Codex integration, plus low, high and max reasoning-effort settings. DeepSeek recommends the high setting for routine agent work.

What are DeepSeek’s new V4-Pro API prices?

DeepSeek’s announced schedule applies two rates, with off-peak usage priced at half the peak level. Peak periods are 1 a.m. to 4 a.m. UTC and 6 a.m. to 10 a.m. UTC. All figures below are per million tokens:

• Input: $0.435 previously, rising to $0.66 off-peak and $1.32 at peak.
• Output: $0.87 previously, rising to $1.98 off-peak and $3.96 at peak.
• Cache hits: $0.003625 previously, rising to $0.022 off-peak and $0.044 at peak.

Against the prior listed V4-Pro rates, off-peak input pricing is about 52% higher and output about 128% higher. Cache-read pricing rises about 507%. At the new rates, a cache hit costs roughly one-thirtieth of regular input pricing, compared with about one-hundred-twentieth previously. That makes the cache change the key cost issue for workloads with substantial repeated context or repeated-file reads.

OpenRouter also displayed the earlier DeepSeek provider rates of $0.435 for input, $0.87 for output and $0.003625 for cache reads per million tokens. As a routing platform, however, it does not independently establish DeepSeek’s future price schedule.

What is DeepSeek Harness?

Harness v0.1 is open-source agent software released as a Developer Preview under the MIT license. The Decoder reported that it uses the Cordis plugin system, allowing components including tools, sandboxes, sessions and the user interface to be swapped. It maintains a session log covering prompts, tool calls and results; runs can be resumed, branched or replayed.

The agent release is distinct from the V4-Pro model itself. The Decoder reported that DeepSeek had not published weights for the new V4-Pro build, while an April preview version remained available on Hugging Face.

DeepSeek also reported benchmark gains for V4-Pro-0813. Its comparison table put Terminal Bench 2.1 at 87.9, up from 72.1, and DeepSWE at 62.7, up from 12.8. Separately, The Decoder said Artificial Analysis raised V4-Pro’s Intelligence Index score from 45 to 53, level with GLM-5.2 and below the scores it listed for Muse Spark, Qwen 3.8 Max, Kimi K3 and Claude Opus 5.

This story draws on original reporting from The Decoder.