This is still a closed frontier model against open weights. DeepSeek V4 Pro is a 1.6-trillion-parameter mixture-of-experts model with 49B parameters active per token and a 1M-token context window, released under the MIT licence with weights on Hugging Face. On DeepSeek's API it costs $1.32 per million input tokens and $3.96 per million output tokens at peak times, and half that off-peak. GPT-6 Astra, OpenAI's most capable model, costs $10 and $50, reads images as well as text, and is only available as a service.
DeepSeek's lineup moved again in September. V4.1-Flash, a smaller model with image input, costs $0.30 and $1.20 at peak, and DeepSeek says tests by multiple parties put it ahead of V4-Pro. V4.1-Pro has not launched yet. Astra's case rests on computer use, long terminal tasks and document work, where OpenAI's published results lead. If a large share of your spend goes to one frontier model, route a sample of your real prompts through DeepSeek in Omny and compare quality and cost before switching.
Which should you choose?
ChatGPT
GPT-6 Astra · 1M context
Pick it if
you need computer use, image input and polished documents and slides.
DeepSeek
DeepSeek V4 Pro 0423 · 1M context
Pick it if
you want MIT-licensed weights, self-hosting or much lower per-token cost.
Pick ChatGPT with GPT-6 Astra for computer use, image input and polished document work. Pick DeepSeek V4 Pro for open weights, self-hosting and far lower API prices. You can also use both: DeepSeek for volume and Astra for the hardest steps.
Questions
Is DeepSeek as good as ChatGPT?
It depends on the task. When V4 launched in April 2026, DeepSeek described V4-Pro in its maximum reasoning mode as the best open-source model available and said it narrows the gap with closed models on reasoning and agent tasks. GPT-6 Astra is newer and adds image input and computer use. Compare them on your own prompts.
Can I self-host DeepSeek?
Yes. DeepSeek V4 Pro and V4 Flash weights are on Hugging Face under the MIT licence. V4 Pro has 1.6 trillion parameters, so running it yourself takes a large multi-GPU deployment. OpenAI's GPT-6 models are only available as a service.