Deepseek V4 Flash matches OpenAI's GPT-5.6 Luna at 60% lower cost
This digest was compiled by AI from multiple sources — links to the originals are below.
Deepseek released V4 Flash "0731" on Wednesday, an AI model scoring 50 points on the Artificial Analysis Intelligence Index — one point behind OpenAI's GPT-5.6 Luna. The model cuts task costs by roughly 60 percent despite OpenAI's recent 80 percent price cut, driven by a 98 percent cache discount. The launch escalates a price war between Chinese and US AI labs, with Moonshot separately securing 20,000 NVIDIA GPUs to scale training.
Technical Gains
V4 Flash "0731" improves across all tested categories, with the largest leap in agentic tasks. On GDPval, a benchmark for complex office work, its Elo rating jumped from 1,189 to 1,559. The model hallucinates less often and uses 12 percent fewer tokens than its predecessor. Its 284-billion-parameter architecture, with 13 billion active and a one-million-token context window, remains unchanged. The weights are released under an MIT license on Hugging Face.
Price War Acceleration
According to Wccftech, the new Flash model operates at approximately $0.28 per million tokens, undercutting OpenAI's budget offering by about 60 percent per task. Deepseek's 98 percent cache discount, well above the industry standard of 90 percent, is a key factor. The move came hours after OpenAI's 80 percent price cut on its own budget models. In a parallel development, Moonshot, the lab behind Kimi K3, has secured 20,000 new NVIDIA GPUs to expand training capacity, signaling broader Chinese investment to challenge US dominance.
What's Next
OpenAI is expected to respond with further pricing adjustments or new model releases. It remains unclear how long Chinese labs can sustain such aggressive pricing given hardware access constraints and regulatory scrutiny.
2 sources
Deepseek V4 Flash matches OpenAI's GPT-5.6 Luna at 60% lower cost
