NewsBite analysis · AI, Startups
The Decoder
New Deepseek Flash model matches OpenAI's GPT-5.6 Luna at roughly 60 percent lower cost
Deepseek has launched its V4 Flash model, an upgrade that competes closely with OpenAI's GPT-5.6 Luna while costing a striking 60 percent less per task. The new model, launched with significant enhancements, scores higher in performance metrics and utilizes fewer tokens than its predecessor. Notably, it has improved in agentic tasks and reduced hallucinations. This push positions Deepseek as a formidable player in the budget AI model sector, reflecting an ongoing shift towards cost-effective AI solutions.
MonitorConfidence: HighShort term horizon
Published 31 Jul 2026, 18:10 UTC1 min read
Visit original article ↗Lead image — The Decoder
Summary
TL;DR
- New V4 Flash model scores 50 points, besting its predecessor by 10 points.
- Costs approximately 60% less than OpenAI's GPT-5.6 Luna.
- Achieves significant improvements across various performance metrics.
- Utilizes 12% fewer tokens than previous versions.
- Architecture remains consistent with 284 billion total parameters.
Interesting facts
- Deepseek offers model weights under an MIT license on Hugging Face.
- The model's performance on GDPval increased from 1,189 to 1,559 Elo points.
Coverage gaps