DeepSeek-V4-Pro Officially Goes Live — One Score Reveals What It's Really After in the Agent Battlefield
The official version of DeepSeek-V4-Pro launched on August 13, with its DeepSWE benchmark score jumping from 12.8 in the preview version to 62.7. It now natively supports the OpenAI Responses API format, while also previewing a new peak/off-peak pricing scheme coming in August.
From 12.8 to 62.7 — that's the DeepSWE benchmark score DeepSeek itself announced, spanning two versions of the same model: April's preview release and the officially launched DeepSeek-V4-Pro on August 13. The gap between them is nearly fivefold.
At Taipei time on August 13, DeepSeek officially announced via its X account (@deepseek_ai) and API changelog that the GA (general availability) version of V4-Pro is now fully live — marking the official launch of the V4 flagship series following April's preview and July's V4-Flash. It's now available across the official app, web platform, and API, with developers only needing to set the model name to deepseek-v4-pro to switch over. Beyond the massive jump in DeepSWE scores, V4-Pro also scored 87.9 on Terminal Bench 2.1, a result the company directly ties to real-world capability in "production environments and Agent tasks."
Beyond the numbers, what really addresses developer pain points in this update is native support for the OpenAI Responses API format, along with adaptation for Codex. The official one-click configuration script sends a clear message: lowering the technical barrier for projects migrating from other platforms. For teams already invested in the OpenAI ecosystem, it's a signal that you can "swap the engine without tearing apart the car."
Thinking mode has also been broken down further this time, with both V4-Pro and V4-Flash now offering three intensity levels: low, high, and max. The official recommendation is low for simple tasks, high for everyday Agent tasks, and max only for complex computation. The underlying specs remain heavyweight: a Mixture-of-Experts (MoE) architecture with 1.6 trillion total parameters, roughly 49 billion activated parameters, support for a 1-million-token context window, and output of up to approximately 384,000 tokens.
Pricing strategy was also previewed. For now, billing continues under the preview-version standard — about 3 RMB per million input tokens and 6 RMB per million output tokens, with cache hits priced lower. The new scheme will take effect at midnight Beijing time on August 17, 2026, introducing "peak/off-peak pricing," where API calls during off-peak hours cost only half the peak-hour rate. Users looking to try it early can currently access it directly through Expert Mode on the official website or via the official app.
