DeepSeek's latest Pro model, V4 Pro 0813, is now available via API only, according to a link post by Simon Willison published on 12th August 2026. Willison wrote that he linked to OpenRouter because DeepSeek did not appear to have any obvious announcement page for the new model. Willison said he had not been able to confirm whether DeepSeek plans to release the open weights, but noted that weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731, which he said made an open-weights release seem likely. In testing the model, Willison reported that he received 'very different looking pelicans' across the model's three reasoning levels — low, medium, and high — a difference he said he had not noticed from any other model. On benchmarks, Willison wrote that figures were released to the Official DeepSeek WeChat Group, then copied into a Reddit post that moderators deleted for being 'low-effort', and subsequently reproduced in an ASCII-art table on Hacker News. No official benchmark documentation was cited in the post.
- DeepSeek V4 Pro 0813 is available via API only, per Simon Willison (12 Aug 2026)
- No obvious official announcement page; linked via OpenRouter
- Open weights not confirmed, though prior V4 Pro and V4 Flash weights are available
- The model has three reasoning levels: low, medium, and high
- Benchmarks circulated informally via WeChat, Reddit, and Hacker House, not official docs
What it means for you
DeepSeek has quietly released a new version of its Pro model, available through API access. For most people this is a minor incremental update — a new model from a Chinese AI lab that competes on price with the bigger names. There is no verified benchmark or official documentation yet, so treat performance claims with caution.
Try this
If you already use DeepSeek or route models through OpenRouter, you can test V4 Pro 0813 on a low-stakes task and compare its three reasoning levels against your current model before switching anything important.
Who should care
Developers and technical teams who already use DeepSeek via API or comparison-shop models on OpenRouter for cost reasons.
Skip this if
You use a hosted assistant like ChatGPT, Claude, or Copilot and don't pick your own models — this changes nothing for you.
Sources: Simon Willison — read the original