DeepSeek-V4-Flash "0731" Nears GPT-5.6 Luna at Roughly 60% Lower Cost
DeepSeek has released the "0731" update, a major upgrade to its low-cost model "V4-Flash". Its score on an AI analysis firm's index climbed 10 points to 50, coming within just one point of OpenAI's "GPT-5.6 Luna", while cost per task is reportedly kept about 60% lower.
"V4-Flash-0731" is the official release replacing the preview version, and its agent capabilities are said to have been significantly enhanced. On Hugging Face, quantized versions (GGUF) from Unsloth and bartowski have been published one after another, and efforts to run the model locally are gaining momentum. Local community evaluations suggest that on coding benchmarks it "beats Opus 4.8 and trails only Opus 5".
The trend of chasing peak performance with less compute and lower cost is now unmistakable — a development that could shake up the pricing of frontier models themselves.
Snapchat Excludes AI-Generated Videos from Monetization, While AI Slop "Pays" on X
Platforms are beginning to respond to the flood of AI-generated content, better known as AI slop. Snapchat says it will adjust its recommendation system so that fully AI-generated videos are excluded from Spotlight recommendations and payouts. By keeping only human-made content, the company is taking a clear stance against AI slop.
Meanwhile, on X (formerly Twitter), AI-generated melodrama videos where "good triumphs over evil" are racking up millions of views and paying their creators. The question for each platform is becoming not whether to ban the technology itself, but whether low-quality auto-generated content gets rewarded with money and algorithmic reach.
Apple Explores Selling "Compute" for Siri AI
Apple CEO Tim Cook has revealed a plan to let users buy additional processing power for a more advanced Siri AI. The move suggests the company would sell extra compute resources through its existing iCloud+ subscription.
Until now, advanced Siri features have been tied to devices and OS versions. But a new pattern is coming into view: selling compute load as a subscription for "AI power users" who run heavy inference in the cloud. Amazon and Google are also pushing harder on offering AI features through paid tiers, and it increasingly looks like big tech's revenue model is gradually shifting from "selling features" to "selling compute".
OpenAI Takes Down a Cambodia-Based Scam Group
OpenAI has disclosed that a Cambodia-based scam group was abusing ChatGPT, and has moved to shut it down. The group reportedly used the tool to support investment solicitation, romance scams, gambling, and impersonation.
Incidents of AI tools being used to make crime more efficient keep surfacing, but what stands out this time is that the vendor itself took the lead on the takedown. Monitoring for abuse and countering it is settling in as a responsibility on par with model development.
xAI's Colossus Data Center: Unpermitted Turbine Removal Drags On
Trouble continues to drag on around the power infrastructure supporting xAI's massive "Colossus" data center. While SpaceX is making progress on a new power plant built for Colossus, the removal of the existing unpermitted gas turbines is expected to be delayed by another several months to a year.
AI's compute demand is running into physical constraints in the form of electricity, and friction over permitting and environmental impact looks likely to shape the pace of future expansion.
EU Commits Up to €30 Billion to AI Gigafactories; US Tech Is Spending 20x That
The European Commission is moving ahead with a plan to build up to seven AI gigafactories across the region. It assumes around €30 billion in combined public and private funding, whereas major US tech companies alone plan to pour more than $600 billion into compute infrastructure this year.
The scale gap is glaring. For Europe to maintain its own AI competitiveness, what will be tested is a comprehensive strategy — not just the amount of funding, but also procurement sources and regulatory design.