AI Agents' Autonomous Task Performance Growing Faster Than Predicted
According to ITmedia, the amount of time AI agents can autonomously work on tasks is growing faster than research institutions predicted. It turns out that the latest models, such as Anthropic's Claude Mythos Preview and OpenAI's GPT-5.5, substantially outperform existing models.
One of the factors that had hindered the practical adoption of AI agents was the limited ability to handle tasks consistently over long periods. The latest generation of models, however, appears to be breaking through this barrier. Longer autonomous runtimes translate directly into practical utility across a wide range of areas, including software development, data analysis, and business automation.
That said, it's worth noting that these performance gains do not show up evenly across all use cases. In tasks involving complex reasoning or scenarios that require integration with external tools, accuracy is still reported to be inconsistent.
Microsoft Edge Copilot Gets a Major Upgrade
Microsoft has announced a major upgrade to Copilot, the AI chatbot built into the Edge browser. The headline feature is the ability to load all open tabs at once. Previously, you had to interact with each tab individually, but the new Copilot can compare and summarize information across multiple tabs.
A long-term memory feature has also been added, making it possible to maintain context across past conversations. There is also a new feature that turns tab content into an AI podcast, along with a quiz mode. While information processing in the browser becomes significantly more efficient, privacy concerns over the AI reading the contents of every tab are being debated in some quarters.
AI writing assistance for LinkedIn has been integrated as well, making Microsoft's strategy of positioning the browser as an AI hub clearer than ever.
Anthropic Moves Programmatic Use of Claude Subscriptions to a Separate Budget
According to The Decoder, starting June 15, Anthropic will separate programmatic usage of Claude subscriptions (calls via the SDK or from third-party apps) from the existing subscription allotment. In its place, users will receive dedicated credits ranging from $20 to $200 per month depending on the plan.
Until now, subscription holders consumed the same allotment whether they used the chat UI or the SDK. Going forward, however, requests made via the SDK or third-party apps will be billed at standard API rates. In effect, the pricing advantage for programmatic usage is coming to an end.
This change could raise costs for developers who use tools like Claude Code on a daily basis. Anthropic, for its part, says the aim is to make API usage more transparent while minimizing the impact on chat UI users.
Three Quarters of AI Customer Service Deployments Fall Short of Expectations and Get Rolled Back
According to The Register, at roughly 74% of companies, AI customer service deployments have underperformed expectations, forcing partial or complete rollbacks. The topic is also drawing attention on Hacker News.
Automating customer support with AI chatbots is an area many companies have pursued over the past few years. In practice, however, it is not uncommon for the effort to backfire: inadequate handling of complex inquiries, growing customer frustration, and ultimately more escalations to human operators.
This shows that AI is not a cure-all, and that human operators remain especially valuable for emotionally charged problem-solving and exceptional situations. The importance of identifying where AI truly fits has now been backed up by real-world data.
Meta: Record Profits, but Internal Morale at an All-Time Low
According to an investigative report by WIRED, Meta is posting record-high profits while internal morale has sunk to its lowest level ever. A workforce reduction of roughly 10% is scheduled for next week, and WIRED interviewed more than ten current and former employees.
Multiple employees echoed the sentiment that "everyone is miserable," and the organizational restructuring that has accompanied massive AI investments appears to be taking a psychological toll. Even with strong profits, a structural problem has come into focus: those profits are not translating into employee motivation or a sense of stability.
Linear CEO: "The AI Honeymoon Is Over"
Karri Saarinen, co-founder and CEO of the project management tool Linear, has been sharing candid views on the current state of AI via Qiita. He points out that it has been about six months since models last made a major leap in coding capability.
He describes a "six-month reality check," in which the initial wow of AI coding is giving way to visible limitations in day-to-day work. The sentiment that productivity gains are real, yet the transformation has fallen short of what was initially hoped for, appears to be resonating within the developer community as well.
Japan's Financial Services Agency Forms a Public-Private Working Group on Mythos-Level AI Cyberattack Countermeasures
According to ITmedia, Japan's Financial Services Agency has convened a public-private working group to strengthen security measures in the financial sector, in response to the improved cyberattack capabilities of the latest AI models, including Anthropic's Claude Mythos Preview. Anthropic's Japanese entity is also participating.
As AI models grow more sophisticated, the risk of their misuse for security attacks is increasing as well. Given how much financial infrastructure affects society, strengthening defenses in advance can be considered a reasonable response. It is worth noting that national-level AI security responses are beginning to take concrete shape.