DeepSeek said on May 23 that it would make its 75% introductory discount on V4-Pro API pricing permanent after the promotion expires on May 31, cementing a rate first scheduled to end on May 5 and later extended. The Chinese AI startup offered no reason, intensifying the cost-of-intelligence race among frontier-model providers.
The new per-million-token prices—$0.003625 for cache-hit input, $0.435 for cache-miss input, and $0.87 for output—sit at exactly one-quarter of the levels published at the April 24 launch. For developers building long-context or agentic applications that burn through millions of tokens, the permanent reduction removes a significant cost hurdle. A footnote on DeepSeek’s English pricing page states that after May 31 at 15:59 UTC, the pricing “will be officially adjusted to 1/4 of the original price.”
The page shows strike-throughs revealing the prior rates: $0.0145, $1.74, and $3.48. The Chinese-language pricing page displays the equivalent yuan figures: 0.025, 3, and 6 yuan per million tokens, down from 0.1, 12, and 24 yuan. DeepSeek did not publish the original Saturday statement cited by Reuters, which first reported the permanence. A universal caveat on the page—“Product prices may vary. DeepSeek reserves the right to adjust product prices”—makes clear that “permanent” denotes the new official list price, not an irreversible guarantee.
V4-Pro is the heavyweight member of the V4 series, carrying 1.6 trillion total parameters with 49 billion active, a 1-million-token context window, and a maximum output of 384,000 tokens. Launched alongside the smaller V4-Flash on April 24, both models came with open weights under an MIT license and API access compatible with OpenAI ChatCompletions and Anthropic interfaces. The introductory 75% discount had been slated to expire on May 5; after an extension, the company permanently locked in the cut.
DeepSeek has repeatedly undercut both Western and Chinese rivals on price. Analysts quoted by InfoWorld, Engadget and the South China Morning Post said the permanent move would intensify pricing pressure on OpenAI, Anthropic, Google, and other proprietary API providers. The cut arrives at a moment when API pricing is becoming the primary competitive lever beyond benchmark scores. The reset puts DeepSeek’s list output price below a dollar per million tokens, a level that undercuts many competing frontier models. Reuters noted that the company did not confirm whether increased supply of Huawei Ascend 950 chips contributed to the decision. The business rationale remains undisclosed.
Whether third-party inference providers such as OpenRouter, Together and Fireworks are passing the lower list prices through to their enterprise customers is unclear, limiting the direct benefit for buyers who rely on resellers. Without passthrough, the headline reduction may not immediately flow to end users, and enterprises that contract through cloud marketplaces could see little change. Geopolitical restrictions, data-sovereignty rules and intellectual-property concerns further constrain adoption of Chinese AI models in several key markets. No competitor has yet announced a matching price adjustment. The real-world impact on enterprise spending could lag behind the announcement.
For technology leaders, the new pricing underscores the accelerating commoditization of frontier model access. The gap between headline token costs and the true expense of deploying AI—once compliance, infrastructure and geopolitical filters are factored in—means the lower API rate may not directly lower enterprise services bills. Still, the permanent discount reinforces a pricing trajectory that forces every provider to either cut margins or differentiate sharply on reliability and ecosystem.