At DevDay, OpenAI unveiled Dots, an agent designed for ongoing work, alongside models and developer tools including GPT-6.1 Sol and Ultrafast. AI products are shifting from answering questions to carrying out tasks over time. Claude in Chrome is also now available on all paid Claude plans, with safety checks for browser actions. Beyond competition on model speed and price, what agents can do autonomously—and when they must ask permission has become just as important.
Several companies are trying to turn AI from a one-off assistant into an agent that can handle ongoing work.
OpenAI unveiled Dots, which lets users specify what an agent may do on its own, what requires approval and what it must not do. Latent Space’s AINews column described how it runs in the cloud and connects to apps; Sam Altman also posted an announcement. Rather than a single question-and-answer exchange, it targets ongoing tasks across apps.
Read original →According to Claude’s official blog, Claude in Chrome is now available on all paid Claude plans. It can use a user’s existing login sessions to read webpages, click links and fill out forms, and can carry out some browser actions autonomously. The company says a safety classifier checks each action before it is taken.
Read original →AI Valley reports that Manus 2.0 has updated its agent architecture and added cloud-computer and automation capabilities. The same briefing describes Cue as a personal-agent product in which each agent can have its own email address, phone number, wallet and computer. These product details come from the briefing; the source material did not include links to corresponding official announcements.
Read original →🔍 Analysis: Dots, Claude in Chrome and Cue share a direction: giving AI access to more tools and longer chains of tasks. They differ in where they run and how permissions work. Whether an agent works continuously in the cloud or uses a browser’s existing login sessions, questions about what it can do and when it needs approval become central to the product rather than optional settings.
This round of releases is a competition not only over capability, but also over the time and cost of completing the same task.
Sam Altman announced GPT-6.1 Sol, saying it costs about one-fifth as much as Astra and offers a 95% discount on cached input. Latent Space’s AINews column also summarized the model’s pricing and test scores. Those performance figures are results cited by the company or in reporting; they cannot be taken as representative of every real-world task.
Read original →The description of an official OpenAI video says Ultrafast runs in Codex at up to eight times the speed of Astra Standard and four times that of Astra Fast. Latent Space’s AINews column also mentions the mode and its higher price. The video has no available captions, so this account cites only the public claims in its description, without drawing conclusions about the demonstration or real-world performance.
Read original →AI Valley says Claude Sonnet 5.5 is more than 30% faster than Sonnet 5, with input and output priced at $2 and $10 per million tokens, respectively. The briefing also judges it well suited to routine tasks with clear boundaries. That is the briefing’s assessment of where to use it, not a claim that it outperforms other models on every task.
Read original →🔍 Analysis: Sol, Ultrafast and Sonnet 5.5 highlight different trade-offs: some emphasize lower usage prices, while others charge more for faster responses. For developers, comparing models means looking beyond any single score and distinguishing test results and vendor claims from the time and cost they see on their own tasks.
Another major thread of the announcements is making agents easier to integrate into existing development workflows.
Latent Space’s AINews column reports that Codex has added a cloud environment where tasks can keep running even after a laptop is closed. It also notes CLI updates, including worktrees and /agents. These changes target development tasks that take longer to complete.
Read original →According to Latent Space’s AINews column, OpenAI has launched the Decisions API for fast multiple-choice classification and routing based on text and images. In plain terms, it handles choices such as which workflow should receive a request, rather than asking a model to write a long answer.
Read original →Vercel CEO Guillermo Rauch posted that AI SDK has reached 30 million downloads per week. That is one measure of the tool’s distribution, but download counts do not equal the number of active developers or show how well specific applications work.
Read original →🔍 Analysis: Long-running cloud tasks, an API that selects workflows, and a widely downloaded developer tool address execution, routing and integration in agent workflows, respectively. Platform competition therefore extends beyond the models themselves to whether developers can reliably fit them into existing processes.
When AI can keep running and operate websites, product promises must be considered alongside safety boundaries.
Claude’s official blog explains that webpages, emails or form fields may contain instructions intended to mislead an agent—an attack known as prompt injection. The company says it improved its defenses before expanding access to Claude in Chrome and checks browser actions before carrying them out. That makes safeguards part of the product design; it does not mean the risk has disappeared.
Read original →Box CEO Aaron Levie argues that, for the foreseeable future, the AI industry can address safety and security through shared standards and practices. He also expects more oversight, testing, accountability mechanisms and regulation as capabilities advance. This is his view of a path for governance, not an industry consensus already in place.
Read original →Latent Space’s AINews column reports that OpenAI has revised its plan tiers and added Pro 500. The column argues that the changes reduce the relative value of the former Pro 200 plan and says they have prompted strong backlash. The plan changes and whether they offer good value are separate questions; the latter depends on how much a user actually uses the service.
Read original →🔍 Analysis: The more permission agents have to act, the harder it is to treat misleading instructions, unauthorized actions and changing costs as minor issues. Today’s discussion spans technical safeguards, industry rules and how users experience pricing. Together, these will shape whether people trust AI that works on an ongoing basis.