DeepSeek API Input Cache Price Reduction

DeepSeek's recently released DeepSeek-V4-Pro model API has initiated a 25% discount special offer, with the input (cache hit) price at 0.25 yuan/million Tokens, the input (cache miss) price at 3 yuan/million Tokens, and the output price at 6 yuan/million Tokens, with the promotion ending on May 5th.

According to Open Router data, the latest overseas renowned AI large model GPT-5.5 Pro has a weighted average input price of 30 USD/million Tokens, with an output price of 180 USD/million Tokens, already exceeding 700 times the difference in input price compared to DeepSeek V4 Pro.

The standard edition of GPT-5.5 has a weighted average input price of 5 USD/million Tokens, and an output price of 30 USD/million Tokens, including the output prices of Anthropic Claude Opus series, OpenAI GPT-5.4, and Google Gemini 3.1 Pro series of large models ranging from 12-25 USD, all of which have a significant gap compared to the newly priced DeepSeek V4 Pro.

On April 24th, the Ronald Keung team from Goldman Sachs published a research report, stating that the newly open-sourced V4 model is a continuation of DeepSeek's efficiency-first, open-source approach.

On the technical level, V4 achieves significant cost reduction in long context windows through architectural upgrades and makes a clear bet on Huawei's domestic chips. On the market level, this release has accelerated the intense competition in China's AI model landscape, with programming capabilities, task completion rates, and multimodality becoming the core dividing lines for pricing power.

Goldman Sachs maintains its recommendation rating for cloud computing and data centers sectors, as the continuous improvement in computational cost efficiency will drive the accelerated penetration of AI applications. The dual engines of AI proxy growth on the enterprise side and AI assistants on the consumer side will support the continuous enhancement of cloud service pricing capabilities.

V4 Architecture Upgrade, Supporting Longer Contexts with Less Memory

DeepSeek V4 is released in two versions: Pro and Flash.

The Pro version is of flagship scale, with a parameter count of 1.6 trillion (activated parameters 49 billion); the Flash version is relatively lightweight, with a parameter count of 284 billion (activated parameters 13 billion). Both models support an ultra-long context window of 1 million token units, on par with top US models (SOTA), but with significantly reduced memory and KV cache requirements.

According to the Goldman Sachs report, V4 Pro requires only 27% of the floating-point operations (FLOPs) needed for DeepSeek V3.2 in scenarios with 1 million contexts, and the KV cache occupancy is only 10%; V4 Flash is even more aggressive, reducing FLOPs to 10% and compressing KV cache to 7%.

This leap in efficiency is achieved through three key architectural innovations:

• In terms of hybrid attention mechanisms, V4 introduces a hybrid architecture of Compressed Sparse Attention (CSA) and Heavy Compressed Attention (HCA). CSA performs sequence dimension compression on the KV cache before executing sparse attention calculations, while HCA adopts more aggressive compression while retaining dense attention, significantly reducing the temporary memory required for long inputs.

• Regarding training stability, V4 introduces the mHC mechanism, enhancing the stability of information transmission in multi-layer networks;

• It also uses Muon as the main training optimizer (with some modules retaining AdamW) to adapt to the more complex network