Key facts
- DeepSeek has launched the official V4 Pro AI model.
- The V4 Pro model boasts enhanced agent capabilities.
- Pricing for V4 Pro has been significantly adjusted, with reports of a 75% reduction.
- The model is designed for cost-efficient long-context inference.
- DeepSeek also released a free version of its 'Harness' tool.
China's artificial intelligence company DeepSeek has officially launched its V4 Pro AI model, which features significantly enhanced agent capabilities. Initially, the model was expected to come with a substantially higher price tag, nearly five times that previously announced.
However, subsequent reports indicate DeepSeek has implemented a steep price cut for V4-Pro, reducing its cost by 75% just a month after unveiling the V4 generation. Usage costs have been revised from a previous range of $0.0145 per million tokens (cache hit) to $3.48 per million output tokens, to a new starting price of $0.003625 per million tokens, up to $0.87 per million output tokens. The company stated this permanent price adjustment is due to efficiency gains in long-context inference, with V4-Pro reportedly using a quarter of the compute and a tenth of the memory footprint of its predecessor.
DeepSeek V4 Pro is described as a 1.6T-parameter Mixture-of-Experts model with 49B activated parameters and a 1 million token context. It is available through multiple API providers, with some clustering at a blended price of $2.17 per 1 million tokens. Experts note that the model's competitive pricing and open-weight availability make it a strong contender against Western AI rivals, particularly for enterprises seeking to reduce inference costs for applications like copilots, document review, and code generation.
While DeepSeek V4 Pro closes the performance gap in areas like complex math and reasoning, it still lags behind Western competitors in broader ecosystem adoption, global support, and hyperscaler integrations. The company also released a free version of its 'Harness' tool, which is similar to Anthropic's Claude Code.