DeepSeek V4 Pro finally ships, and the price warning is the real news
DeepSeek's flagship left preview today. V4 Pro 0813 is now the model behind the deepseek-v4-pro endpoint, ending a preview that ran nearly four months. The numbers: 1.6 trillion total parameters, 49 billion active per token, a one-million-token context window, and a maximum output of 384,000 tokens. Hybrid attention, tuned for full-codebase analysis and long-horizon workflows rather than chat.
Pricing carries over from preview. $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, $0.87 per million output. That cache-hit number is not a typo β it is roughly a hundred and twenty times cheaper than a miss, which tells you exactly what DeepSeek thinks agent workloads look like. An agent re-reads the same repo, the same system prompt, the same tool schemas on every single turn. Price the repeat reads at near zero and a hundred-turn agent loop stops being a budget line item.
But the sentence people actually stopped on sits in DeepSeek's own notice: they plan to raise overall API pricing in the near future, with a significant increase expected. A lab that built its entire reputation on being the cheap frontier option is telling you, in advance, that the cheap part is ending. Half the Hacker News thread was developers who moved off GitHub Copilot in April doing quiet math about what happens next.
That is the honest read on today. The model is a solid GA release and the benchmark gains are concentrated in coding and agentic use, which is where DeepSeek has been aiming since V4. The strategic story is that the discount era was customer acquisition, and acquisition is over. DeepSeek now sits second only to Anthropic on token consumption. You do not warn people about price hikes when you are still trying to win them.
The model is live at https://openrouter.ai/deepseek/deepseek-v4-pro-0813 and through DeepSeek's own API.
← Back to all articles
Pricing carries over from preview. $0.435 per million input tokens on a cache miss, $0.003625 on a cache hit, $0.87 per million output. That cache-hit number is not a typo β it is roughly a hundred and twenty times cheaper than a miss, which tells you exactly what DeepSeek thinks agent workloads look like. An agent re-reads the same repo, the same system prompt, the same tool schemas on every single turn. Price the repeat reads at near zero and a hundred-turn agent loop stops being a budget line item.
But the sentence people actually stopped on sits in DeepSeek's own notice: they plan to raise overall API pricing in the near future, with a significant increase expected. A lab that built its entire reputation on being the cheap frontier option is telling you, in advance, that the cheap part is ending. Half the Hacker News thread was developers who moved off GitHub Copilot in April doing quiet math about what happens next.
That is the honest read on today. The model is a solid GA release and the benchmark gains are concentrated in coding and agentic use, which is where DeepSeek has been aiming since V4. The strategic story is that the discount era was customer acquisition, and acquisition is over. DeepSeek now sits second only to Anthropic on token consumption. You do not warn people about price hikes when you are still trying to win them.
The model is live at https://openrouter.ai/deepseek/deepseek-v4-pro-0813 and through DeepSeek's own API.
Comments