DeepSeek V4 Pro 0813 Appears on OpenRouter With No Announcement
DeepSeek quietly released its V4 Pro 0813 model on 12 August 2026, making it available via API through OpenRouter but publishing no official announcement page.

👁 121 reads
DeepSeek released its V4 Pro 0813 model on 12 August 2026, making it accessible via API on OpenRouter but publishing no official announcement page of its own. The release was flagged by developer Simon Willison, who noted he had to link to OpenRouter because DeepSeek provided no direct announcement.
- DeepSeek V4 Pro 0813 is available via API on OpenRouter as of 12 August 2026, with no official announcement from DeepSeek
- Open weights have not been confirmed, though previous models DeepSeek-V4-Pro and DeepSeek-V4-Flash-0731 both had weights released publicly
- Benchmark figures were reportedly shared only in DeepSeek's official WeChat group before spreading informally to Reddit and Hacker News
The model is currently API-only. Whether DeepSeek will release open weights remains unconfirmed, though the pattern with earlier releases points toward that outcome - weights for both April’s DeepSeek-V4-Pro and July’s DeepSeek-V4-Flash-0731 are publicly available on Hugging Face, making a similar release here seem likely.
One technical curiosity Willison noted: the model produces noticeably different outputs depending on whether reasoning is set to low, medium, or high - a variation he described as unlike anything he had observed with other models.
The benchmark situation is unusually murky. According to Willison, performance figures were released exclusively to DeepSeek’s official WeChat group, from where they were copied into a Reddit post that moderators deleted for being “low-effort,” and then reproduced in an ASCII-art table on Hacker News. No benchmark data has been formally published by the company.
For developers and builders working with frontier models, the practical picture here is straightforward but worth noting: a capable model appears to be available on OpenRouter right now, but without official documentation or verified benchmarks, any production decision rests on direct testing rather than published specs. The reasoning-level sensitivity Willison flagged also suggests that default API settings may produce meaningfully different results from those obtained at higher reasoning budgets - worth checking before committing to a prompt design or cost estimate.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →More on this from QUANTSLANT
ai tools·Microsoft merges two Copilot apps into one, ahead of ‘super app’

