DeepSeek released preview versions of DeepSeek-V4-Pro and V4-Flash on April 24, 2026, adding multimodal capability that lets the models process images alongside text. V4-Flash reportedly performs on par with Anthropic’s Claude Opus 4.8 across reasoning, coding and agentic tasks while costing roughly $0.28 per output, compared with about $25 to $30 for comparable work from Anthropic. Its efficiency is linked to using approximately 90 KV cache entries for image processing, versus roughly 870 for Claude models, as well as a Mixture-of-Experts architecture that activates only part of the model for each task. The models support context windows of up to 1 million tokens and are compatible with OpenAI and Anthropic APIs, reducing switching costs for developers. Follow-up releases, including V4-Flash-0731 and the experimental deepseek-v4-flash-vision-exp, indicate rapid iteration. The launch extends the cost-cutting strategy DeepSeek established with its R1 model in January 2025, while reports of distillation attacks have heightened tensions with Western AI companies.