DeepSeek Announces "Substantial" Price Hikes — a Turning Point for Its Low-Price Strategy?
DeepSeek, the Chinese company that has kept US and Chinese rivals at bay with disruptive low prices, has announced that it will implement "substantial" price increases across its AI services. According to the notice the company posted, no specific increase amounts were given — it simply urges users to be prepared.
DeepSeek currently offers its API at some of the lowest prices in the industry, such as $0.14 per million input tokens, making it emblematic of the race to drive down model costs. For DeepSeek itself to raise prices is unusual, and it could force a rethink of the assumption that "inference from China is cheap." Rising compute costs and growing demand appear to be behind the move, though details remain unknown at this point. The news is also drawing attention on Hacker News and may offer a clue to the fate of the API economy that has been sustained by low prices.
Korea's Kakao Publishes the Technical Report for "K-EXAONE 2.0"
A technical report has been released for K-EXAONE 2.0, the new version of "EXAONE," a frontier-class large language model developed by Korea's Kakao. With major US and Chinese models on the rise, the fact that Korean players continue to develop their own frontier models is worth noting in the context of a multipolar AI landscape. As of this writing, we have not been able to review detailed performance evaluations in the report, and we await future benchmark results.
Governments' Bets on the AI Boom Are "Dangerous," The Economist Warns
In an editorial, Britain's The Economist argued that governments around the world are betting too heavily on the AI boom. Under the headline "a dangerous bet," it sounds the alarm over public investment and industrial policy built on the assumption that AI will continue to grow sustainably. The size of the economic benefits AI will bring — and the timing of their arrival — remain uncertain and could prove more limited than assumed, the piece argues, urging that policy assumptions be kept flexible.
"Permission Design" for MCP and Agents Becomes a Practical Challenge
While MCP (Model Context Protocol), which connects models to external services, continues to spread, know-how on permission design is being shared one after another for in-house operations. Some argue that when hosting MCP servers internally, you should work through authentication, roles, and auditing in that order. Others point out that connecting AI agents to Notion, GitHub, and Google Drive via personal OAuth can give them access to "files that were never meant to be visible."
Meanwhile, how to design configuration files when rolling out Claude Code to a team is also sparking debate, with some publishing where they draw the lines between deny, ask, and allow — along with the reasoning behind those choices. As agents operate ever more autonomously, minimizing the permissions granted and monitoring them closely seems to be becoming part of the development process itself.
You Can't Learn by "Dumping Everything on AI" — the Quality of Engagement Is Key
Several empirical studies have been reported that back up that end-of-day feeling of not being able to remember the code you wrote after a day of AI-assisted development. In randomized controlled trials (RCTs), handing learning-oriented tasks outright to AI reduced how much participants learned, while those who engaged in dialogue with the AI and verified its output achieved results on par with or better than the group that did not use AI at all.
The issue, in other words, is not whether to use AI but how to engage with it. Now that generation speed has increased, there may be room to rethink the quality of the human side of that engagement.