DeepSeek's V4-Flash outperforms premium models with 645% gain

DeepSeek just redefined the agentic AI landscape with the public beta release of its DeepSeek-V4-Flash API (model ID: deepseek-v4-flash), a lightweight model that outperforms its premium V4-Pro-Preview counterpart across every agent benchmark despite identical architecture. The breakthrough comes from 'better post-training'—the same model size but radically improved agentic capabilities. On Terminal Bench 2.1, V4-Flash scored 82.7 versus V4-Pro-Preview's 72.1, a 14.7% lead, while costing just $0.14 per million input tokens (cache-miss) compared to Anthropic's Opus-4.8, which scores 85.0 on the same benchmark but at significantly higher pricing.

The model's performance gains are staggering: a 645% improvement on DeepSWE (from 7.3 to 54.4) and dominant scores across NL2Repo (54.2), Cybergym (76.7), and Toolathlon Verified (70.3). DeepSeek also natively supports the OpenAI Responses API format and Codex integration, making it a seamless drop-in replacement for OpenAI models in agent pipelines. This positions V4-Flash as the most cost-effective frontier agent model, enabling teams to slash costs by up to 90% while maintaining high performance.