DeepSeek V4 Pro Reaches GA with Adjustable Reasoning and Responses API Support

DeepSeek first released the V4 preview on 24 April 2026. On 13 August, the company announced the general availability build now identified in its documentation as DeepSeek-V4-Pro-0813.

The API model name remains deepseek-v4-pro, so existing callers do not need to switch to a dated model ID. DeepSeek says the update improves agent performance in production, but that is a vendor claim until independent evaluations test the GA build on comparable workloads.

The GA release adds adjustable reasoning effort to V4 Pro and V4 Flash. DeepSeek maps low effort to simpler tasks, high to everyday agent work, and max to more complex jobs.

DeepSeek also added native support for the OpenAI Responses API and published a Codex setup path. This gives agent tools another integration option beyond the OpenAI-compatible chat completions and Anthropic interfaces already documented for V4.

DeepSeek lists V4 Pro in Expert Mode on its app and website, and through the official API as deepseek-v4-pro. The company documentation now points that stable API alias to the 0813 build.

AIMI currently records the exact 0813 route on OpenRouter, Cloudflare Workers AI and NVIDIA NIM. OpenRouter first appeared at 18:01:03 SAST on 12 August, Cloudflare at 20:43:56 SAST on 14 August, and NVIDIA NIM at 00:59:11 SAST on 27 August. These are provider observations, not additional maker release dates.

NVIDIA's official Build page now lists deepseek-ai/deepseek-v4-pro-0813 as a preview endpoint on its NIM service. The page dates the Build listing to 24 August 2026, points to DeepSeek's model card for the underlying release, and identifies the service as a trial endpoint governed by NVIDIA's terms.

AIMI first observed the route in NVIDIA's models endpoint at 22:59:11 UTC on 26 August, or 00:59:11 SAST on 27 August. That was exactly 12 days, 4 hours, 15 minutes and 15 seconds after the Cloudflare route was first observed. The gap describes hosting availability, not a second DeepSeek release or a response by NVIDIA.

DeepSeek introduced separate peak and off-peak API rates at 16:00 UTC on 16 August 2026. The company says off-peak rates are 50% below peak pricing.

The current pricing page lists the active windows and rates. Teams moving recurring work to V4 Pro should check that page before hard-coding figures into cost models.

The April date belongs to the V4 preview. The 13 August date belongs to DeepSeek's GA announcement and 0813 model build, while the official model card records the weights publication on the same calendar day. NVIDIA's Build page gives its hosted listing a separate 24 August date.

DeepSeek's announcement gives a calendar date but no precise publication time. AIMI therefore keeps announcement, weights, provider metadata and endpoint first-seen times separate. There were no earlier verified releases in this package's 27 August SAST batch, so the NVIDIA addition is not being framed as part of a same-day release race.