DeepSeek said on May 23, according to Reuters, that it would make a 75% promotional discount on its V4-Pro API permanent, converting a limited-time offer into the official price after May 31 and heightening price competition among frontier AI model providers as token-heavy workloads strain inference budgets.
The Chinese company’s English-language API pricing page now states: "After the 75% off promotion, the DeepSeek-V4-Pro model API pricing will be officially adjusted to 1/4 of the original price starting from 2026/05/31 15:59 UTC." The new permanent rates are $0.003625 per million cache-hit input tokens, $0.435 for cache-miss input and $0.87 for output, down from $0.0145, $1.74 and $3.48. The Chinese-language page lists equivalent yuan rates of 0.025, 3 and 6 per million tokens, down from 0.1, 12 and 24. Pricing for the smaller V4-Flash model, launched alongside V4-Pro, remains unchanged at $0.0028, $0.14 and $0.28. The cut arrives as developers increasingly run agentic workloads that consume tens of thousands of tokens per session, making inference cost a decisive line item.
V4-Pro is the larger member of DeepSeek’s V4 family, introduced on April 24. It uses 1.6 trillion total parameters with 49 billion active parameters and supports a one-million-token context window and a 384,000-token maximum output length. The promotional 75% discount had already been in place by April 27, according to Bloomberg.
DeepSeek has repeatedly used price as a competitive lever. In February 2025, Reuters reported it offered off-peak discounts of up to 75% for its R1 and V3 models. On April 26 it cut cache-hit input prices across all models to one-tenth of launch levels. The latest move turns a launch promotion into a permanent rate, adding to an industry-wide price war as coding agents, document analysis and other token-hungry applications spread.
DeepSeek’s pricing page states that the official adjustment will take effect after the promotion expires on May 31 at 15:59 UTC; the Chinese counterpart lists an equivalent end time of 23:59 Beijing time. The company’s terms reserve the right to change prices, so the new rate is not a binding guarantee. But by calling it the official post-promotion price, DeepSeek signals that it intends to keep the lower level for the foreseeable future.
For developers and startups, a permanently lower V4-Pro rate removes a deadline that could have pushed planning toward alternatives. The absence of published user migration data means the competitive impact remains inferred. Still, the move signals that DeepSeek aims to sustain pricing pressure, and budget-constrained teams may factor the lower rate into build-versus-buy decisions for agentic features, where per-token costs compound quickly.
Open questions persist. It is not known what capacity constraints, rate limits or latency trade-offs apply at the discounted price, nor whether major third-party API distributors such as OpenRouter have updated pass-through pricing. DeepSeek did not disclose what drove the decision; Reuters noted that the company did not link it to expanded supply of Huawei Ascend 950 chips. Competitor responses—from OpenAI, Anthropic, Google and others—had yet to surface by May 23. While the lower price is attractive, few independent benchmarks compare V4-Pro’s reasoning quality, latency and uptime against Western models, leaving buyers to weigh cost against reliability for now.
The promotional period ends on May 31. From that date, the market will watch whether the pricing holds, how resellers adjust, and whether rival labs cut list prices or offer their own concessions. In the meantime, the adjustment cements DeepSeek’s position as the most aggressive price-setter in the frontier-model market, a stance that could compel competitors to respond before the end of the third quarter.