DeepSeek V4-Pro Launch Cuts AI Token Costs vs GPT-5.5 Pro
DeepSeek has released V4-Pro (1.6T total parameters, 49B active) and V4-Flash with 1M-token context windows. DeepSeek V4-Pro targets a mix-of-experts design and long-context efficiency using Compressed Sparse Attention and Heavily Compressed Attention, cutting compute and KV-cache costs versus V3.2.
Pricing is the headline for builders and AI agents: DeepSeek V4-Pro costs about $1.74 per million input tokens and $3.48 per million output tokens, while V4-Flash is about $0.14/$0.28. The article frames this as roughly 1/20th the cost of Claude Opus 4.7 and 98% cheaper than GPT-5.5 Pro. DeepSeek also claims that once additional “supernodes” come online later in 2026, the Pro model’s price can drop further.
On capabilities, DeepSeek reports strong coding and agentic performance. It highlights “deep reasoning” modes (V4-Pro-Max) and introduces “interleaved thinking” so multi-step tool-using agents retain reasoning context across tool calls.
Strategically, the release is positioned within the US–China AI chip-export restrictions context, with DeepSeek reportedly training partly on Huawei Ascend hardware to work around export limits. The timing follows OpenAI’s GPT-5.5 launch, adding to competitive pressure in the AI model market—an area that can indirectly influence crypto sentiment around AI infrastructure and risk appetite.
Neutral
该消息对加密市场的直接影响有限，但存在情绪与叙事层面的“间接效应”。DeepSeek V4-Pro 主打成本大幅下降（98%低于 GPT-5.5 Pro、约 1/20 低于 Claude Opus 4.7）与更高效的长上下文/Agent 能力，这通常会强化“AI算力与模型效率提升=行业竞争加速”的预期，从而短期内提升风险偏好、利好与 AI 基建叙事相关的资产。
不过，历史上类似“模型/算力成本下探”更多改变的是科技与产业预期，并不必然立刻传导到链上资金流或宏观流动性；真正推动加密价格波动通常还依赖比特币/美元流动性、利率预期、监管与链上交易数据等。综合来看，市场可能对“AI基础设施投资逻辑”做轻微再定价，但幅度大概率受限，因此定性为 neutral。
短期（几天到一两周）：更像是板块轮动与情绪扰动。
长期（几个月）：如果更低成本能带动AI应用放量，才可能通过算力/生态投资间接形成更稳定的叙事支撑。