DeepSeek Flash is equal to GPT-5.6 Luna at 40% the cost

DeepSeek released their latest update, V4 Flash ‘0731’. This model achieves a score of 50 on the Artificial Analysis Intelligence Index, while GPT-5.6 Luna by OpenAI gets a score of 51. It’s 60% cheaper, with 284B parameters but only 13B active, 1M context size, and MIT license.

GDPval showed that this model went from 1,189 to 1,559 Elo. DeepSeek also gives a 98% discount on cached tokens (instead of 90%), which makes it particularly appealing for agentic tasks that require repetition.

Chinese bargain-basement models have caught up with US flagship models to within a single benchmark point. However, they’re still 2x cheaper. This is putting tremendous pressure on the AI industry as a whole, as more and more businesses struggle to afford expensive API calls.