NewsBite analysis · AI, Business, Startups
The Decoder
Thomas Joos
Deepseek has launched its upgraded AI model, V4 Flash '0731', which claims to match OpenAI's GPT-5.6 Luna in performance while being 60% cheaper per task. This new model demonstrates a significant improvement, scoring 50 points on the Artificial Analysis Intelligence Index, just shy of GPT-5.6 Luna. It boasts substantial gains in efficiency and lower hallucination rates, making it a compelling choice for businesses looking for cost-effective AI solutions.
Act nowConfidence: HighShort term horizon
Published 31 Jul 2026, 18:10 UTC1 min read
Visit original article ↗Lead image — The Decoder
Summary
TL;DR
- Deepseek's V4 Flash '0731' now scores 50 points, surpassing its prior version.
- It costs 60% less per task compared to OpenAI's GPT-5.6 Luna.
- The model uses 12% fewer tokens than its predecessor.
- Improvements noted in agentic task performance.
- The model architecture includes 284 billion parameters with a million-token context window.
- Weights available under MIT license on Hugging Face.
Interesting facts
- Deepseek's pricing strategy is significantly lower than industry standards.
- The model shows vast improvements in real-world office work benchmarks.