In the early hours of August 13, 2026 (KST), DeepSeek quietly swapped the model version in its API docs to DeepSeek-V4-Pro-0813. There was no separate announcement or blog post.
The preview period, which ran for almost four months, has ended and the model moved to general availability (GA). It is the first official release since the open-weight V4 family came out on April 24. (GMI Cloud)
The news spread through user posts before official channels. The account @ChrisGPT posted the comparison table DeepSeek released and wrote:
"Terminal-Bench 2.1 went from 72.1% to 87.9%, CyberGym from 52.7% to 83.3%, and DeepSWE from 12.8% to 62.7%." (2026-08-13 01:12 KST)
Up to 5x vs. the Preview, but the Baseline Was Low
In the table DeepSeek released, the biggest gain is on DeepSWE. It jumped 4.9 times, from the preview's 12.8% to the release's 62.7%.
| Benchmark | V4-Pro-0813 | V4-Pro-Preview | Kimi-K3 | Opus-4.8 | Fable 5 |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 87.9 | 72.1 | 88.3 | 85.0 | 88.0 |
| Cybergym | 83.3 | 52.7 | 80.0 | 78.3 | 83.1 |
| DeepSWE | 62.7 | 12.8 | 67.5 | 58.0 | 70.0 |
| AutomationBench (Public) | 31.8 | 12.8 | 30.8 | 27.2 | 29.1 |
Source: DeepSeek's official announcement image (via the X post by @ChrisGPT, checked 2026-08-13 01:12 KST)
Pricing is $0.435 per million input tokens and $0.87 per million output tokens. The context window is 1 million tokens. (Based on the @ChrisGPT post)
Next to Kimi-K3 and Fable 5, It's Not First
In the same table, Kimi-K3 (Terminal-Bench 88.3, DeepSWE 67.5) and Fable 5 (Terminal-Bench 88.0, DeepSWE 70.0) still score higher than DeepSeek V4 Pro.
Compared with Opus-4.8, DeepSeek V4 Pro leads on most metrics. The accurate summary is "not at the very top, but greatly improved over the preview and close to the top tier."
V4-Pro Performance Is Self-Reported Only
All of these figures are materials DeepSeek released itself. Results reproduced by an independent outside organization have not been confirmed yet. Agent benchmark scores can swing widely depending on harness setup, prompts, and tool environment, so caution is needed.
DeepSeek's official blog and changelog pages also do not yet have a separate post announcing this release. For now, the API price list and user reports are the most concrete primary information.