Google has introduced Gemini 4 Argon for coding, knowledge work, and cyber defense, but access is limited for now. Another major theme at OpenAI DevDay is enabling AI agents to work continuously and take part in real workflows through computer use and developer APIs. Google Labs and Vercel are also making it easier to handle recurring tasks and connect services. As agents take on more tasks, permissions, credentials, and the risk of operational mistakes demand closer attention.
Today’s model news is about more than capabilities: access and real-world performance matter too.
Google has introduced Gemini 4 Argon for coding, enterprise knowledge work, and cyber defense; both Latent Space’s AINews column and AI Valley covered the launch. Google says it ranks first on 13 of the 19 benchmarks it published, but initial access is limited to select government users and trusted cyber defenders. Its maximum 1 million output tokens requires Long Decode Continuation, which pauses and resumes a long response across multiple calls. That is not the same as the output limit for a standard single call.
Read original →OpenAI’s Thibault Sottiaux said demand for GPT-6.1 Sol is high across both the API and subscriptions, putting a heavy load on ChatGPT and Codex. He said the team had added capacity and expected speeds to improve over the next few hours, potentially reaching nearly twice the previous day’s speed. That was an expectation at the time, not a measured result.
Read original →🔍 Analysis: Argon stands out for its reported capabilities and exceptionally long output, but limited access means most people cannot yet test it themselves. The news about Sol is a reminder that reliable, fast delivery also shapes the experience of using a model. When comparing models, it helps to distinguish vendor-reported benchmark results from the conditions under which features work and whether the model is actually available.
From persistent agents to computer use, the focus of new releases is shifting from answering questions to completing tasks.
AI Valley covered Dots, a continuously running AI agent introduced by OpenAI at DevDay. It has a cloud computer and can ask the user to step in when needed. OpenAI also released a Dots demo video today; its title and description indicate only that it shows scenarios involving travel planning, organizing user feedback, and following up in Slack. No captions were available for the video, so its specific steps or results cannot be assessed from that information.
Read original →OpenAI’s Thibault Sottiaux said users can now build and deploy MCP servers directly through ChatGPT, then restrict who can access them or make them public. MCP can be understood as an interface convention for connecting AI to external tools. This update puts interface creation and access settings in the same workflow.
Read original →Latent Space interviewed Ari Weinstein, who works on computer use at OpenAI, about how agents combine screenshots, accessibility information, web page structure, and generated code to operate software, and how they recover from failures. The point is not that an agent merely “looks at images and clicks buttons,” but that it can use several kinds of information to determine the state of an interface. Assessments of speed and capability in the interview remain the interviewee’s claims.
Read original →In the same Latent Space interview, Nikunj Handa of OpenAI’s API team discussed asynchronous tool calls, changing instructions during a run, WebSockets, and the Decisions API. In plain terms, these capabilities concern letting agents avoid waiting continuously for tools to return, accept new instructions while a task is underway, and communicate with applications more promptly. The interview also covered caching and context compaction for long-running tasks.
Read original →🔍 Analysis: Together, these updates sketch out a task workflow: an agent needs to keep running, understand and operate software, and connect to external tools, while developer APIs help the whole process respond more promptly. As capabilities expand, the scope of access and the way permissions are granted become part of product design. Whether an agent finishes a task is not the only question.
Bringing AI into everyday work is becoming a product battleground of its own.
Google Labs said Skills in Gemini is now available globally, letting users save custom instructions in chats and automate recurring tasks. Google Labs linked it to its earlier Opal experiment in workflow customization. The announcement concerns a feature rollout and provides no quantified results on time saved.
Read original →Vercel CEO Guillermo Rauch invited service providers to join Connect to make it easier for applications and agents to access services. He argued that the challenge in building applications is shifting toward “connecting existing services” and pointed to the difficulties of giving every agent its own static key. This is Rauch’s account of the product direction, not independent verification of Connect’s security benefits.
Read original →🔍 Analysis: Skills is about reusing a person’s instructions and steps; Connect is about linking agents to external services. They address different points in a workflow: first making a task repeatable, then giving it access to the systems it needs. The broader those connections become, the more important credentials and access controls are.
How to indicate the origin of AI-generated content remains a technical issue worth watching.
Google DeepMind released a video in which Hannah Fry discusses AI watermarking with Pushmeet Kohli and Jeremy Ratcliffe. Its title and description indicate that the topics include SynthID and watermarking applications in text, images, video, and biology. Because captions were unavailable, this account does not infer the video’s specific arguments, experimental results, or conclusions.
Read original →Tools are advancing quickly, but views still differ on how people’s work and business processes will change.
Box CEO Aaron Levie sees a significant opportunity in deploying AI across enterprise economic activity because changing business workflows often takes more work than expected. He said this could take the form of software and agents, as well as generate demand for new services. This is his assessment of the market and the challenges of deployment, not an established industry conclusion.
Read original →Investor Nikunj Kothari said he has automated most of his work, but still personally finds and contacts entrepreneurs, meets founders, and writes explanations for decisions not to invest. He specifically said he writes each of those explanations himself. This is an account of his own work, not a boundary on automation that can be applied to every role.
Read original →