DeepSeek-V3 ends promotional pricing, updates API service rates

DeepSeek has officially concluded its 45-day introductory promotional pricing for its flagship DeepSeek-V3 model, transitioning enterprise users and developers to updated standard API service rates. While the company's aggressive pricing structure helped drive rapid global adoption following its launch, the shift to standard commercial tiers is reshaping downstream enterprise economics—prompting the first major client departure due to operational cost concerns.

Quick Facts

• Promotional End Date: DeepSeek-V3's 45-day introductory pricing period ended on February 8, with revised rates taking effect on February 9.
• New Input Pricing: RMB 0.5 ($0.068) per million tokens for context cache hits; RMB 2 ($0.27) per million tokens for cache misses.
• New Output Pricing: RMB 8 ($1.09) per million output tokens generated.
• First Major Exit: AI infrastructure firm Luchen Technology announced on March 1 that it will drop DeepSeek API services within a week over financial concerns.
• Profitability Paradox: The client exit follows recent disclosures from DeepSeek claiming a theoretical profit margin of 545% for its online inference systems.

What Happened

DeepSeek-V3, the open-weight large language model that created waves across the tech industry with its high efficiency and low resource footprint, has formally wrapped up its initial promotional phase. Beginning February 9, all developer and enterprise accounts utilizing the DeepSeek API were moved to the provider's standard price matrix.

The conclusion of the 45-day discount marks a pivotal moment for DeepSeek as it pivots from aggressive user acquisition toward long-term commercial monetization. However, the price normalization has already generated friction within the artificial intelligence ecosystem, culminating in the first high-profile enterprise customer dropping the service.

Key Details

Under the finalized fee schedule, DeepSeek has implemented a differential pricing model that penalizes un-cached queries while rewarding efficient context management. The detailed breakdown of the updated rates includes:

• Cache Hit Input: RMB 0.5 (approximately $0.068) per million tokens when queries hit existing prompt caches.
• Cache Miss Input: RMB 2 (approximately $0.27) per million tokens for un-cached or newly processed input context.
• Output Generation: RMB 8 (approximately $0.09 to $1.09 range, specifically benchmarked at $1.09) per million output tokens.

Although these rates remain competitive compared to standard commercial models offered by Western hyperscalers, the relative jump from promotional tiers significantly alters unit economics for third-party platforms building directly on top of DeepSeek's API.

Background

DeepSeek-V3 gained widespread industry attention due to its novel Mixture-of-Experts (MoE) architecture and Multi-head Latent Attention (MLA) framework, which drastically reduced training and inference overhead. To encourage immediate developer integration upon release, DeepSeek instituted a 45-day discounted API rate.

The economic efficiency of DeepSeek's infrastructure made headlines when the company reported a theoretical profit margin of 545% for its online operations. This internal efficiency demonstrated that DeepSeek could generate substantial margins even at low price points. However, high operational margins for the foundational provider do not necessarily translate into affordable, predictable costs for downstream integrators facing high query volumes.

Why It Matters

The real-world financial impact of the updated pricing became clear on March 1, when AI infrastructure provider Luchen Technology announced it would officially discontinue its support for the DeepSeek API. Luchen Technology represents the first major technology firm to formally drop DeepSeek services citing direct cost