Pre-IPO Profit Push! DeepSeek Raises V4 Model Prices Over 4x Amid IPO Preparation

DeepSeek is aggressively boosting profitability ahead of a planned initial public offering (IPO) by sharply increasing prices for its flagship V4 AI model. However, even after the price surge, the China-based AI company—known for its low-cost offerings—still charges significantly less than key rivals.

According to an announcement posted Thursday (August 13) on DeepSeek's official website, a new peak-hour pricing structure will raise certain rates by over four times current levels. The revised pricing takes effect on August 16.

Under the new model, the DeepSeek-V4-Flash will cost $1.32 per million output tokens during peak hours and half that during off-peak periods—up from the previous rate of just $0.28 per million tokens. A token is a unit measuring the computational workload required for an AI model to process information and generate results.

This price adjustment signals DeepSeek’s intensified focus on profitability as it pursues a public listing. Yet, its per-token cost remains substantially lower than that of leading competitors. For instance, Moonshot’s Kimi K3 charges $15 per million output tokens, while Anthropic’s most advanced Fable 5 service costs $50.

The V4-Pro model will also see a significant increase: $3.96 per million tokens during peak hours and half during off-peak, up from the current $0.87.

The Hangzhou-based AI lab stated the pricing revision aims to “more reasonably allocate resources.” This dynamic pricing strategy is designed to incentivize developers and enterprises to shift workloads to less congested network periods. Last week, DeepSeek had issued a preliminary warning about price increases but did not disclose specific figures.

DeepSeek’s pricing strategy has recently sparked a global debate on the cost of high-performance AI tools. Its historically low pricing—so aggressive it created concerns about a so-called 'death zone' where higher-cost or weaker models are eliminated—has raised questions about the IPO plans of Western counterparts like OpenAI and Anthropic.

Bloomberg previously reported that DeepSeek is in a major fundraising phase and has begun preparations for a potential IPO as early as this year. Founder Liang Wenfeng now faces the challenge of balancing investor expectations, rapid market expansion, and massive capital requirements for building expensive computing infrastructure.

On the same day, DeepSeek also unveiled the developer preview of DeepSeek Harness v0.1, signaling its intent to compete with services like Anthropic’s Claude Code.

A 'harness' refers to a software framework built around large language models (LLMs), enabling AI agents to handle complex tasks such as reading files, editing code, browsing the web, and persistently executing workflows until completion.

Unlike some U.S. competitors who hard-code product features, DeepSeek emphasizes an open architecture that allows users to integrate any desired components—including other AI models—to strengthen its AI agent ecosystem.