DeepSeek has officially ignited a massive “AI price war” by announcing a 75% discount on its newly released flagship model, DeepSeek-V4-Pro, on Monday, April 27, 2026.
This aggressive promotional period runs until May 5, 2026, and positions DeepSeek as the most cost-effective frontier-level intelligence provider in the world, significantly undercutting Western rivals like OpenAI, Anthropic, and Google.
1. The 75% Discount Breakdown
The promotional pricing applies specifically to the DeepSeek-V4-Pro API, which features a massive 1.6 trillion parameter Mixture-of-Experts (MoE) architecture.
| Token Type | Standard Price (per 1M) | Discounted Price (per 1M) |
| Input (Cache Miss) | $1.74 | $0.435 |
| Input (Cache Hit) | $0.145 | $0.036 |
| Output | $3.48 | $0.87 |
- Comparison: At $0.87 per million output tokens, DeepSeek-V4-Pro is roughly 34x cheaper than OpenAI’s GPT-5.5 ($30/1M) and 28x cheaper than Anthropic’s Claude 4.7 Opus ($25/1M).
2. The “Deep-Water Bomb”: 90% Cache Reduction
Alongside the V4-Pro discount, DeepSeek launched a permanent structural change to its entire API suite (including the V4-Flash variant) by reducing the cost of input cache hits to one-tenth of their original price.
- The Logic: This specifically targets AI Agents that frequently send similar or repeated long-context requests (like a coding agent reading the same codebase repeatedly).
- Result for Flash: Before May 5, the input cache hit price for DeepSeek-V4-Flash has dropped to an astonishing $0.0029 per million tokens (approximately ¥0.02).
3. Strategic Rationale
Industry analysts view this as a “dimensional reduction strike” aimed at forcing developers to switch ecosystems.
- Incentivizing Switchers: By making the world’s top-tier intelligence virtually “too cheap to ignore,” DeepSeek aims to capture the developer market for agentic frameworks like OpenClaw and Claude Code.
- Domestic Chip Optimization: The V4 series is natively optimized for Huawei Ascend 950 chips. This move proves that high-performance AI can be deployed at scale without relying on Nvidia hardware, a critical pivot for the Chinese tech stack amidst export restrictions.
- Market Pressure: Even affordable providers like Alibaba Cloud (Qwen3) are suddenly finding their pricing uncompetitive compared to the V4-Pro promotional rates.
4. Technical Specs: DeepSeek-V4 Series
Released in preview on April 24, 2026, the V4 series represents the first major architecture update since December.
- Context Window: 1 million tokens for both Pro and Flash.
- Architecture: Mixture-of-Experts (MoE).
- Pro: 1.6T total / 49B active parameters.
- Flash: 284B total / 13B active parameters.
- License: MIT (Open Weights).