What is confirmed is that DeepSeek has indeed updated its V4 Pro model for the first time since April, and just recently posted this message to its website:
«🎉 The official version of DeepSeek-V4-Pro has been released, featuring significantly enhanced agent capabilities and support for the Responses API and Codex integration. It is now fully available across the web, mobile app, and API; we welcome your testing and feedback.»
The kicker lies in the price, which ticks in at $0.435 per 1M input tokens and $0.87 for 1M outputs. This pricing might not last, as DeepSeek posted a warning of significantly higher costs just days ago.
The official Pro version performs better than Gemini 3.6 Flash and GPT-5.6 Luna on the official Artificial Analysis benchmark of benchmarks.
Second best open weights model
Artificial Analysis’ Intelligence Index has it as one of the best open weights models at 53 points, falling behind only Kimi K3, which scores 60 points.
It sits slightly below Muse Spark 1.2, which costs $1.25/M in and $4.25/M out and scores 57, but on par with GLM-5.2, which costs $1.4/M in and $4.4/M out — both priced at around three times as much.
There have been no other officially verifiable benchmarks posted so far. DeepSeek V4 Flash is 18th on the Arena.ai benchmark, with V4 Pro expected to score significantly better.
The DeepSeek V4 Pro official release is a 1.6 trillion parameter model with a one million token context length, and you can find it here or wait for the weights on Hugging Face.
Read more: DeepSeek English with announcement, API pricing, Artificial Analysis rundownm South China Morning Post. Discussion on Hacker News and r/Singularity.