DeepSeek slashes AI model prices in push to undercut rivals

DeepSeek sharpened the economics of the AI race by making its newest model dramatically cheaper to use. The Chinese startup offered a 75% discount on DeepSeek-V4-Pro until May 5 at 15:59 UTC, while cutting input cache-hit prices across its API lineup to one-tenth of the launch price.

That matters because pricing has become a deciding factor for developers and startups deciding which model to build on. DeepSeek’s pricing page shows V4-Pro priced in per-1M-token terms and spells out discounted cache-hit, cache-miss and output rates during the promotion, a structure designed to encourage experimentation while pulling in traffic for a fresh release. The immediate message to the market is clear: adoption comes first, margins later.

The discount comes on the heels of DeepSeek’s preview launch of its V4 model on April 24, when the company said the new open-source system arrived in pro and flash versions and improved knowledge, reasoning and agentic capabilities. That agent focus is strategically important. Companies are racing to build systems that can do more than answer questions, including longer workflows that require tool use, coordination and more compute.

DeepSeek said the V4 Pro version outperformed other open-source models in world-knowledge benchmarks and trailed only Google’s closed-source Gemini-Pro-3.1. If those claims hold up in broader use, the combination of stronger performance and sharply lower access costs could force rivals, including U.S. model makers, to rethink how they price premium access for developers.

The launch also showed how closely the pricing war is tied to hardware strategy in China. DeepSeek adapted the new model for Huawei chip technology, underscoring Beijing’s push for tech autonomy as it tries to reduce dependence on foreign systems and supply chains. The V4 rollout follows DeepSeek’s late-2024 V3 release and builds on the low-cost R1 model that had already rattled Silicon Valley and world markets.

Industry analysts had expected the new model to arrive more than a month earlier, around the start of the Lunar New Year, but DeepSeek’s timing now looks calibrated to keep momentum on its side. By pressing prices down so aggressively on a fresh model, the company is also signaling that the path to profit in generative AI may get narrower before it gets wider, especially as global competitors face higher compute costs, faster model cycles and growing pressure to prove that premium pricing still makes sense.