DeepSeek's new model is 105x cheaper than rivals: Here's why

DeepSeek is back doing what made it famous in the first place: undercutting everyone else on price by an enormous margin.

Research firm Artificial Analysis found that DeepSeek's newly released V4-Flash model costs just 3 cents on average per benchmark test, against $3.15 for Anthropic's Claude Fable 5, $1.86 for OpenAI's GPT-5.6 Sol, and 86 cents for Moonshot AI's Kimi K3.

The model charges $0.14 per million input tokens and $0.28 per million output tokens, putting it roughly 105 times cheaper to run than Fable 5 on the firm's tests.

Just because the headline price is low does not guarantee a bargain. The comparison made by Artificial Analysis takes into account the amount of data which the martificial analysisnd produces in order to finish a certain task, because the model which requires many more steps in order to get to the answer may cost more in total even though its quote is lower.

On Artificial Analysis's Intelligence Index, which consists of nine tests, including coding, reasoning, and workplace-related tasks, V4-Flash managed to get only 50 out of 100 points, making it equal to Google's Gemini 3.6 Flash but one point behind Meta's Muse Spark 1.1 and Z.AI's GLM-5.2 and significantly behind Moonshot's Kimi K3, which got 57 points.

DeepSeek’s R1 model caused a sell-off in the technology stocks around the globe at the beginning of 2025 with evidence of the possibility of ultra-low-cost AI, but now, the company has been crowded out by domestic competitors including Moonshot, MiniMax, and Z.AI, as well as ByteDance and Alibaba.

DeepSeek has apparently lined up an IPO on its home turf with a valuation of $74 billion. The launch of V4-Flash seems like a last ditch effort by DeepSeek to reclaim its position of being synonymous with low cost inlast-ditch crowding space.