From Temporary Promotion to Permanent DeepSeek API Pricing Reset
DeepSeek has converted what looked like a time-limited promotion into a structural reset of its API pricing. The company is making the 75% discount on its V4-Pro model permanent, locking prices at one-quarter of the original launch rate after the discount period ends on May 31, 2026 at 15:59 UTC. The updated price card lists DeepSeek-V4-Pro at USD 0.435 (approx. RM2.01) per million uncached input tokens and USD 0.87 (approx. RM4.02) per million output tokens, down from crossed-out reference prices of USD 1.74 (approx. RM8.04) and USD 3.48 (approx. RM16.07). Cached input now costs USD 0.003625 (approx. RM0.02) per million tokens. The cheaper DeepSeek-V4-Flash tier comes in at USD 0.14 (approx. RM0.65) per million input tokens, USD 0.28 (approx. RM1.29) per million output tokens, and USD 0.0028 (approx. RM0.01) for cache hits, reinforcing DeepSeek’s push toward affordable AI development for high-volume workloads.
Why a Permanent API Discount Matters for Developer Budgets
For AI-native startups, DeepSeek’s move is more than a headline-grabbing AI model cost reduction. Many young companies have discovered that their products behave like software, but their cost structure resembles usage-based infrastructure. Every support reply, research summary, coding task, or agent decision generates a token bill that can quietly erode gross margins. By turning a 75% discount into a permanent API discount, DeepSeek gives teams a stable baseline for forecasting costs, especially for long-context workloads that pass large prompts and documents repeatedly. Cheaper token rates and sharply reduced cache-hit pricing across the lineup make it more realistic to keep rich context inside prompts instead of constantly trimming, summarizing, or pushing users toward minimal experiences. That predictability allows founders to test lower-ticket features, expand free tiers, or serve price-sensitive user groups without betting the business on a promotion that might vanish at renewal time.
Huawei’s Ascend Chips and the New Ceiling for AI Model Cost Reduction
DeepSeek has not formally spelled out how it can sustain such aggressive pricing, but industry attention is focusing on Huawei’s Ascend AI chip ecosystem. DeepSeek previously acknowledged that limited access to high-end compute pushed V4-Pro pricing far above its cheaper Flash model at launch, with Pro access reportedly costing up to 12 times more due to constrained advanced hardware. Now, usage for V4-Pro is quoted between 0.025 and 6 yuan per million tokens, down from a prior range of 0.1 to 24 yuan, suggesting that inference costs are dropping as hardware supply improves. Huawei’s Ascend 950 chips have become increasingly important where alternative advanced accelerators are restricted, potentially providing the capacity DeepSeek needs to trade margin for reach. If this hardware-backed efficiency holds, the company’s pricing could force a broader recalibration of what “normal” AI inference costs look like across the industry.
Competitive Shock: How DeepSeek’s Pricing Pressures Rivals and Platform Choices
DeepSeek’s locked-in pricing puts sharp pressure on rivals and forces a new kind of feature–price trade-off for developers choosing a platform. Analysis of rate cards suggests that, for some workloads, V4-Pro can be 20 to 35 times cheaper than premium offerings from leading Western providers, depending on prompt structure and output volume. That gap will be hard for budget-conscious teams to ignore, particularly for agents, coding assistants, and retrieval-heavy document systems that churn through millions of tokens per day. The cut also lands in a market already drifting downward on API costs, pushing competitors such as Kimi, Qwen, and MiniMax into tougher comparisons whenever teams run spreadsheet-style evaluations. Yet price is not everything: reliability, latency, data handling, tool support, and trust profiles will still determine which workloads migrate. Nonetheless, DeepSeek has reset expectations for what a high-end, long-context model should cost in sustained production use.
What This Means for Startups Designing With Long-Context AI
For founders and developers, DeepSeek’s new pricing structure changes how products can be architected from day one. Long-context models like V4-Pro and V4-Flash, paired with very low cache-hit costs, make it more practical to persist large instruction blocks, documents, and knowledge bases directly in prompts. That reduces the engineering overhead of aggressive summarization layers or complex prompt-rotation schemes built solely to save tokens. Startups can experiment with richer AI-driven experiences in education, small-business tools, and consumer apps where ticket sizes are low and margins were previously squeezed. At the same time, teams must weigh strategic exposure to a single provider, evaluate latency and regional access constraints, and consider whether certain sensitive workflows should stay on alternative platforms. Still, as DeepSeek’s permanent API pricing flows into budgets and financial models, the default assumption that advanced AI is inherently expensive is likely to weaken, unlocking new product categories and business models.