A version of Chinese start-up DeepSeek’s (深度求索) flagship artificial intelligence (AI) model is by far the least expensive to run on benchmark tests among well-known models globally and more than 100 times cheaper to run than Anthropic PBC’s Claude Fable 5, San Francisco-based Artificial Analysis said. DeepSeek’s R1 model became a global sensation early last year, triggering a sell-off in global technology stocks and raising questions about the large amounts US companies were spending on AI. The start-up officially released its V4-Flash model on Friday, its latest attempt to regain momentum by doing what it is best known for — offering ultra-low-cost AI alternatives.
A person holds a cellphone displaying the DeepSeek logo in an illustration photograph taken on Jan. 29 last year.
The V4-Flash charges US$0.14 per million input tokens and US$0.28 per million output tokens, Artificial Analysis said. A token is a unit of data used to measure AI usage.
The research firm estimated V4-Flash’s average cost at US$0.03 per test, compared with US$0.86 for Kimi K3 from Chinese rival Moonshot AI Technology Co (月之暗面), US$1.86 for OpenAI’s GPT-5.6 Sol and US$3.15 for Claude Fable 5. The comparison provides a more realistic measure of value than pricing alone because it accounts for the amount of data a model must process and generate to complete a task. A model with low headline price can still prove expensive if it requires significantly more steps to produce an answer. DeepSeek once commanded most of the headlines about Chinese AI development, but was quickly besieged by many domestic rivals including other start-ups such as Moonshot, MiniMax Group Inc (稀宇科技) and Z.AI Co (智譜), as well as tech giants like ByteDance Ltd (字節跳動) and Alibaba Group Holding Ltd (阿里巴巴). All are vying with US tech firms for global adoption, targeting businesses seeking cheaper ways to deploy AI at scale. Artificial Analysis said DeepSeek’s V4-Flash model scored 50 out of 100 on its Intelligence Index, which combines results from nine benchmarks spanning coding, reasoning and workplace-style assignments. That is the same score as Google’s Gemini 3.6 Flash, and one point behind Meta Platforms Inc’s Muse Spark 1.1 and GLM-5.2 from Z.AI, which is also known as Zhipu. However, Moonshot’s Kimi K3 scored a 57 while Anthropic’s Claude Opus 5, Fable 5, and OpenAI GPT-5.6 scored nine or more points higher. DeepSeek is also preparing a more powerful version of its model, called V4-Pro. It has not given a date for that version’s official release. Separately, Alibaba yesterday unveiled its largest and most capable AI model to date, Qwen3.8-Max, which is not far behind in size when compared with an offering from domestic rival Moonshot AI launched last month.
