DeepSeek updated its cheap model and it now runs neck and neck with OpenAI's cheap model
V4 Flash 0731 gained ten points on the Artificial Analysis index without changing size or price, landing one point behind GPT-5.6 Luna at roughly 60 percent lower cost per task.
DeepSeek has shipped an update to its budget model, and the jump is larger than the modest name suggests. V4 Flash “0731” scores 50 on the Artificial Analysis Intelligence Index, ten points above the version released in April, and six points above DeepSeek’s own more expensive V4 Pro. It lands one point behind GPT-5.6 Luna, OpenAI’s cheapest model, which OpenAI had just cut by 80 percent the day before.
What makes this notable is what did not change. Same architecture, 284 billion total parameters with 13 billion active. Same one-million-token context window, meaning the amount of text it can hold in mind at once. Same price. The gains came from training, not from a bigger or costlier model. It improves across every tested category, with the biggest jump in agentic tasks: on GDPval, a benchmark built around realistic office work, it climbs from 1,189 to 1,559 Elo points. It also hallucinates less than its predecessor and uses about 12 percent fewer tokens to get the same work done.
On cost per task, DeepSeek comes in roughly 60 percent below GPT-5.6 Luna even after OpenAI’s cut. A large part of that gap is not the headline price at all but the cache discount. When you send the same text repeatedly, for example a long system prompt, providers charge less for the repeated part. The industry standard discount is around 90 percent. DeepSeek’s is 98 percent, and at high volume that compounds. The weights are on Hugging Face under an MIT licence, which allows unrestricted commercial use and modification.
What is behind this. This is the price war landing where the money actually is. Flagship models get the headlines, but most real AI spending goes on high-volume routine work: classification, extraction, summarising, moderation. That tier is now genuinely competitive, with DeepSeek, Gemini 3.6 Flash and GPT-5.6 Luna all clustered within a point or two of each other. OpenAI cut Luna by 80 percent on July 30 and DeepSeek matched the moment the next day, which tells you how tight this segment has become. The pressure is good for buyers and uncomfortable for anyone financing large data centres on the assumption of fat margins.
What this means for you: if you pay for AI through a monthly subscription, this does not touch your bill. If you build anything on an API, it is worth an hour of your time. The practical pattern is tiered routing: send the boring high-volume work to a cheap model like this one, and reserve a frontier model for the genuinely hard requests. Teams doing that are cutting bills substantially without losing quality where it counts. A fair caveat before you switch: DeepSeek is a Chinese provider, so if data residency or regulatory approval matters in your organisation, check that first. The open weights under MIT are one answer to that, since you can host the model yourself.
Sources
Europe opens bidding for seven AI gigafactories, and the numbers explain the problem
The European Commission wants up to seven large AI compute sites, backed by 10 billion euros of public money meant to pull in 20 billion more. US tech firms will spend twenty times that this year.
