Token Photo: VCG

A wave of price cuts of tokens has swept through the artificial intelligence (AI) large model sector recently in China and overseas, following a period of price hiking earlier this year. Industry analysts said that the token price reductions reflect substantial cost savings from AI-related technological advances, intensifying market competition and the industry faces reshaped commercialization routes.

A token is the smallest unit of data that AI large models use for processing. It is created by breaking down text or other content into smaller pieces, such as words, sub-words, or characters.

In China, token price cuts began earlier than elsewhere. As one of the latest examples, Volcano Engine, ByteDance's AI services division, on Tuesday launched the Seedance 2.0 mini video AI generation model, with image-to-video pricing at merely 0.023 yuan ($0.004) per 1,000 tokens and video-to-video at 0.014 yuan per 1,000 tokens - a significant reduction compared to the previous version.

And, Tencent Cloud announced that, starting Monday, it would cut prices for the MiniMax-M3 model on its tokenHub platform by 50 percent for inference input, output, and cache hit fees.

Earlier, DeepSeek in late May cut token prices for its V4-Pro model by 75 percent, with input cache price dropping to 0.025 yuan ($0.004) per million tokens, while the price of DeepSeek-V4-Flash was as low as 0.02 yuan per million tokens. 

An industry insider told the Global Times that DeepSeek's current tokenpricing is approaching the electricity cost required to generate the tokens.

Xiaomi followed DeepSeek by reducing prices for its MiMo-V2.5 series by up to 99 percent. 

Internationally, OpenAI took the lead by slashing prices for its o3 reasoning model by up to 80 percent. Input tokens now cost $2 per million, while output tokens are priced at $8 per million. 

"Currently, token prices are influenced by essentially two key factors: the rapid technological progress, namely the technological capabilities of the AI models, and hardware-side cost control. Both areas are now advancing rapidly," a technician from a Beijing-based AI company told the Global Times on condition of anonymity.

Advances in several factors, such as multi-level caching and optimization for domestically-made chips, have significantly lowered inference costs. The technician cited DeepSeek's V4 series, saying that it reduced computing power consumption to 27 percent of its previous generation.

"At the hardware level, falling chip and electricity costs - aided by China's 'East Data, West Computing' initiative - have supported lower pricing," said the technician.

In addition to improvements in electricity infrastructure, new computing-power infrastructure has emerged, which will further reduce the cost of AI-related products. 

For example, the world's first prefabricated computing power hub has been launched in Qingdao, East China's Shandong Province, on June 7, providing a faster and more cost-effective solution for building data centers, according to a CCTV News report.

Data is known as the "feedstock" for the development of AI. The hub can reduce overall costs by 20 percent. It can be completed in about five months, ensuring a stable power supply for a computing center while cutting construction costs by nearly 80 percent, the report said.

Pan Helin, a member of the Information and Communication Economy Expert Committee of the Ministry of Industry and Information Technology, told the Global Times that construction is carried out using prefabricated modules that are later integrated on site. 

"The approach greatly shortens the construction cycle, reduces costs, and speeds up computing power deployment. It creates a mature investment model that helps ease the current token supply-demand imbalance," said Pan.

Lower token prices have translated into strong user growth. According to OpenRouter data, in the first week of June, Chinese AI models recorded 14.19 trillion tokens in weekly usage, far surpassing US models' 3.2 trillion. Chinese models have dominated global usage rankings for six consecutive weeks.

Industry insiders believe that this wave of token price cuts marks a critical turning point as the sector moves from explosive growth toward high-quality development. As more AI companies approach capital markets, user scale and real-world effectiveness are becoming increasingly important.

"Looking ahead, low token prices are likely to become the norm, with large AI models evolving into essential digital infrastructure similar to electricity or water. The current price cut is the beginning of a new phase where only companies that excel in technology, product, and business model innovation will survive and thrive," Wang Peng, an associate research fellow at the Beijing Academy of Social Sciences, told the Global Times.