DeepSeek V4 Pro exits preview at a higher price

DeepSeek moved its flagship V4 Pro model out of preview on August 13, nearly four months after previewing the V4 series on April 24. The GA build, DeepSeek-V4-Pro-0813, is a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters, a 1-million-token context window, and up to 384K output tokens, and it now natively supports OpenAI’s Responses API format with Codex integration and three reasoning-effort levels (low, high, max). DeepSeek reports large agentic gains over the preview build: Terminal Bench 2.1 up to 87.9, DeepSWE up to 62.7, CyberGym up to 83.3.

The catch is price. Current API pricing is $0.435 per million input tokens (cache miss) and $0.87 per million output, but DeepSeek’s own changelog confirms a pricing change effective August 16 introducing peak and off-peak rates, with off-peak priced at half of peak. Calling the model is unchanged, developers still target the deepseek-v4-pro endpoint to get the latest build automatically.

What it means for you

The benchmark jump is real and worth testing if you’re running agentic coding workloads, but don’t assume the preview economics that made DeepSeek attractive still hold after August 16, run your own volume through both peak and off-peak windows before you commit budget. This follows the same pattern as DeepSeek’s DSpark inference speedup earlier this year: DeepSeek ships real capability gains, then adjusts pricing once the preview crowd has proven out demand, so re-price your workload at every graduation from preview to GA rather than assuming the number you budgeted for is the number you’ll pay.