The Vibe Father
DeepSeek V4 Pro GA is the general release, not another preview rumor. DeepSeek’s August 13 changelog says the GA build is on the app, the web product, and the API. The calling method does not change. Set the model name to deepseek-v4-pro and you get the latest version, labeled DeepSeek-V4-Pro-0813 on the pricing page.
That is the event the July watch posts were waiting for. Preview V4 Pro has been in the wild since April. This update is a post-training jump aimed at agents, plus a native Responses API path that DeepSeek says is adapted for Codex.
What changed in the GA notes
- The lab says agent quality is much stronger in production environments.
- Thinking effort on V4-Pro and V4-Flash is now low, high, or max.
- The API supports the OpenAI Responses format and a one-click Codex setup script.
- Peak and off-peak prices take effect at 16.00 UTC on August 16, 2026.
Context stays at one million tokens. Maximum output is 384,000 tokens. Concurrency on Pro is listed at 500, against 2,500 on Flash. If you run a wide agent farm, that cap is part of the product, not a footnote.
What you should do this week
| If you already call deepseek-v4-pro | Do this |
|---|---|
| Pin behavior with golden tasks | Re-run the same repository suite against 0813 before you trust the new traces. |
| Use thinking mode | Map jobs to low, high, and max instead of leaving max on overnight. |
| Pay the API bill | Model the August 16 peak sheet now. Off-peak is half of peak. |
| Care about open weights | Confirm whether the GA weights you need are the ones already on Hugging Face, or a later drop. |
The live Vibe Bench card for this build is DeepSeek V4 Pro 0813. It is a measured profile, not the vendor score table. Keep those two lists separate when you brief a team.
Bottom line
DeepSeek V4 Pro GA is live under the same name you already know. Update your notes, re-run your suite, and put the August 16 price change on the calendar. Do not treat a silent model-name upgrade as a no-op.
