DeepSeek Makes the 75% Discount Permanent: What $0.87/M Output Tokens Changes
Agentic AIDeepSeek’s 75% promotional discount — originally scheduled to expire May 31 — is now the permanent price. Output tokens are $0.87 per million, at least 34x cheaper than GPT-5.5 at $30 per million, and up to 51.7x cheaper on long-context workloads.
The Pricing Decision
The original promotion read as a customer acquisition move with an expiration date. Making it permanent is a different signal: DeepSeek is pricing to become infrastructure, not to trial. At $0.87/M output, the price is low enough that cost is no longer the primary reason to stay on a more expensive model — only capability gaps are.
The 34x figure is the comparison that matters for most AI applications. GPT-5.5’s $30/M output pricing reflects frontier capability positioning. DeepSeek’s pricing reflects a different thesis: many production use cases don’t need frontier capability, and the market for “good enough at scale” is larger than the market for best-in-class.
The Routing Implications
The practical effect of permanent DeepSeek pricing is not that teams switch wholesale from frontier models. It is that the economic case for model routing becomes compelling at a much lower engineering investment. Routing simple queries, classification tasks, summarization, and templated generation to DeepSeek at $0.87/M while reserving complex reasoning for frontier models was already a reasonable architecture. With the discount permanent, there is no longer a planning timeline risk — teams can build routing logic against a price floor that will not reset in six weeks.
The 51.7x long-context pricing gap matters specifically for document processing, multi-turn conversation, and RAG pipelines where context window size drives cost. XAI’s Grok 4.3 price cuts follow a similar competitive dynamic: the frontier model pricing tier is compressing under pressure from Chinese model providers, and the models positioned as “premium” in 2025 are being re-tiered toward commodity pricing in 2026.
The question this raises for teams who invested in DeepSeek V4 Pro evaluations earlier this year: the benchmark case for V4 Pro was partly about price-to-performance. That calculus has not changed — it has been locked in.
The So What
The strategic shift is away from “which model is best” and toward “which tasks actually require the best model.” Permanent $0.87/M pricing from a production-capable model gives teams a durable lower bound to route against. Build the routing layer now, before frontier pricing inevitably compresses further. The firms that will lead on AI unit economics in 2026 are the ones that treat model selection as an ongoing operational decision rather than a one-time architecture choice.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →