Agents Stay On. Models Cost Less. Robots Keep Going.

What happened
OpenAI launched Dots, persistent agents powered by GPT-6 Astra that get their own cloud computer, can work toward goals around the clock, and can connect to more than 4,000 apps through OpenAI’s plugin ecosystem. Users can interact with a dot through ChatGPT, Slack, or Microsoft Teams, while OpenAI is initially rolling the product out across eligible Pro and Business Premium users, with enterprise deployments controlled by administrators.

Why it matters
Dots move the agent model beyond “ask, answer, stop”: the agent retains working context, operates across applications, can pursue several projects, and asks for approval around sensitive actions. That makes continuity, permissions, supervision, and identity management core product features rather than add-ons to an LLM.

What’s next
OpenAI says it is already piloting specialist dots for organizations, including agents with separate identities and access controls, and plans integration with Microsoft Agent 365; its longer-term vision is teams of dots rather than one primary agent per user.

What happened
IBM and Indian cloud provider Yotta announced general availability of a Sovereign Agentic AI Platform for Indian organizations, combining IBM watsonx Orchestrate with Yotta’s Shakti Cloud and Shakti Studio. The stack is designed to keep AI data, inference, operations, and governance within India and is available through Yotta regions in Panvel and Greater Noida.

Why it matters
“Sovereign AI” is expanding from where data is stored to where agents run and how they are governed. For enterprises that cannot casually send sensitive workflows into a global cloud, a locally operated agentic control plane creates a more credible path from AI experimentation to production automation.

What’s next
IBM and Yotta are targeting workflows including security operations, document processing, and HR automation. The real test is whether sovereignty becomes a deployment advantage for agents in regulated and data-sensitive industries rather than simply another infrastructure requirement.

What happened
OpenAI introduced GPT-6.1 Sol, priced at $2 per million input tokens and $10 per million output tokens, and says its own evaluations show performance close to GPT-6 Astra on agentic coding, computer use, and professional tasks at roughly one-fifth of Astra’s standard token price. Sol is available through ChatGPT Work, Codex, and the API but not regular Chat while cached input costs $0.10 per million tokens.

Why it matters
For long-running agents, model economics compound: every planning step, tool call, retry, and retained context consumes inference. If production performance resembles OpenAI’s evaluations, Sol substantially lowers the cost floor for keeping capable agents running continuously instead of reserving frontier intelligence for only the hardest requests.

What’s next
OpenAI says an Ultrafast Sol option is coming to Codex in the following days, with token generation as much as eight times faster. That points to a new competition layer where developers increasingly choose not just by intelligence, but by the combined economics of intelligence, latency, and sustained agent runtime.

What happened
The White House launched America.gov, an AI-powered gateway for federal services spanning information from roughly 29,000 government websites. Google is the technology partner in the initative leveraging Gemini and said the service is intended to help more than 100 million people access government resources; same-day reporting also identified xAI’s Grok as part of the AI stack.

Why it matters
This is generative AI moving from an internal productivity tool to a citizen-facing service layer at national scale. When an LLM mediates access to passports, benefits, loans, healthcare information, and other public services, grounding, reliability, provenance, and error handling become operational requirements with real-world consequences.

What’s next
The bigger shift is planned for early 2027: America.gov is expected to expand beyond answering questions toward completing forms, applications, and renewals. That would turn the system from an AI search interface into something much closer to a transactional government agent.

What happened
Dyna Robotics unveiled Dyna-2.1 alongside its semi-humanoid Taku robot, framing the system as a physical agent that can complete end-to-end workflows rather than one predefined manipulation task. In its launch demonstration, Dyna showed what it described as an uncut, hour-long autonomous commercial-laundry workflow involving multiple interconnected steps.

Why it matters
Physical AI is starting to be judged on a tougher metric: how long a robot can keep useful work going without human intervention, not whether it can nail one impressive pick-and-place demo. Workflow-level autonomy better captures the reliability, recovery, sequencing, and environmental reasoning needed for robots to deliver economic value outside a lab.

What’s next
Dyna says its next milestone is taking the system into real customer sites and building a deployment-data flywheel in which operating experience improves subsequent versions. The benchmark now shifts from demonstration quality to whether those hour-long runs survive messy, repetitive, real-world operations at scale.

The center of gravity in AI is moving from model IQ to operational leverage: agents that stay on, models cheap enough to run continuously, infrastructure built to satisfy sovereignty constraints, and robots measured by how long they can keep working without intervention. The durable advantage increasingly belongs not to whoever has the smartest standalone model, but to whoever makes autonomy economical, governable, and dependable across software and the physical world.

Give an agent a job that doesn’t end when the chat does.

Pick one recurring task you normally revisit every day: checking a dashboard, reviewing new leads, monitoring a shared inbox, or summarizing updates from a project.

Set it up in OpenAI Dots with the apps it needs, then define three things:

Goal: what outcome it should keep working toward.
Boundary: what it can do without asking you.
Escalation: what should always require approval.

Then run the same workflow once with GPT-6.1 Sol and compare the economics of keeping it active versus using a frontier model for every step.

The question is no longer “Can the agent do it?” It’s “Can you leave it running?”