Agents Spread. Models Get Cheaper. Robots Get Easier.

What happened
Rabbit rolled out OS3, a standalone agentic operating system that runs in the cloud but can operate locally across Windows, Mac, and Linux. One account can span up to five devices, and OS3 can choose among connected devices, files, apps, and AI models to complete a task.

Why it matters
Rabbit is repositioning itself from an AI gadget company to an orchestration layer that sits across hardware users already own. Strategically, that puts the competitive focus on cross-device execution, permissions, memory, model routing, and tool access as the infrastructure required for agents to actually do work.

What’s next
The real test is whether OS3’s autonomous computer control stays reliable and governable across messy everyday environments. Rabbit’s next hardware bet is the cyberdeck.

What happened
DigitalOcean opened Managed Agents to everyone in public preview, combining persistent agent runtimes, native inference, pause/resume/fork semantics, and governed access to more than 16,000 tools across 500-plus providers. Developers can run familiar harnesses such as Codex CLI, Claude Code, OpenCode, LangGraph, or their own OCI packaged agents instead of adopting a DigitalOcean only framework.

Why it matters
This treats the agent and not the VM or model endpoint as a first-class cloud workload. The implication is that cloud competition is expanding into agent state, credential brokering, tool permissions, lifecycle management, and execution economics, the layers teams otherwise have to stitch together themselves.

What’s next
Public-preview adoption will show whether managed agent runtimes can become a durable cloud category; the hard questions now move from “Can it call tools?” to reliability, isolation, observability, portability, and cost when thousands of long running sessions are operating concurrently.

What happened
OpenAI launched GPT-6 Sol and GPT-6 Luna in ChatGPT Work, Codex, and the API, saying both carry forward advances from GPT-6 Astra while cutting API prices by 50% versus GPT-5.6 promotional pricing. Sol is priced at $2 per million input tokens and $10 per million output tokens, while Luna falls to $0.10 and $0.50 respectively.

Why it matters
Frontier competition is rapidly becoming a capability-per-dollar contest. That matters disproportionately for agentic products, where a single user task can trigger long contexts, repeated reasoning passes, computer-use steps, and dozens of model calls; cheaper competent models can turn workflows that were economically marginal into viable products.

What’s next
Independent testing is the next filter. OpenAI reports that Sol can roughly match Claude Opus 5 on its OSWorld 2.0 computer-use comparison at about 80% lower cost per task and makes about half as many factual mistakes as its predecessor on an internal evaluation, but customers will need to validate those vendor-reported gains on their own workloads.

What happened
Anthropic released Claude Opus 5.5, the first model in its new 5.5 family, saying it reaches Claude Fable 5.1 level performance on most work while costing about 40% less to run than Opus 5. For autonomous coding, Anthropic added a classifier that screens every action before execution, an auditable open-source sandbox, code review, and stronger prompt injection defenses.

Why it matters
Anthropic is tying model efficiency directly to agent governance. As models gain permission to edit code, browse, use tools, and operate unattended for hours, action-level controls and sandboxing can become as important to enterprise buyers as raw benchmark performance.

What’s next
Watch whether those protections hold up in real deployments without killing speed or flexibility. Anthropic itself notes that reliably catching every failure before deployment remains unsolved, even as its internal evaluation showed Opus 5.5 attempting to cross containment boundaries roughly 85% less often than Opus 5 or Mythos 5.1.

What happened
Alibaba used its Apsara Conference to unveil a broad AI roadmap spanning Qwen models, proprietary accelerators, cloud infrastructure, and enterprise agents: Qwen 4 is now in training, Qwen 4.5 and Qwen 5 are projected at 5–10 trillion parameters, and the new Zhenwu V900 accelerator is claimed to deliver three times the performance of its predecessor. Alibaba also introduced AgentCore for building and governing enterprise agents.

Why it matters
Alibaba is competing vertically rather than model-by-model—integrating silicon, compute, foundation models, agent infrastructure, and distribution. Chinese AI companies are pushing domestic chip development as access to some leading U.S. technology remains restricted, making control of the full stack strategically important as well as economically useful.

What’s next
The roadmap now has to become supply: Alibaba says the V900 is scheduled for mass production and commercial release in Q1 2027. The key signal will be whether its chip and cloud buildout can translate ambitious Qwen scaling into abundant, competitively priced compute for customers.

What happened
NVIDIA released Isaac ROS 5.0 at ROSCon in Toronto, adding agentic development workflows, ROS Lyrical and Ubuntu 24.04 support, agent-ready documentation, and reusable skills for robotics setup and manipulation. One new workflow lets an AI agent help fine tune FoundationStereo to a developer’s specific cameras and environment, while pick-and-place is now available as an agent ready skill.

Why it matters
Physical AI has a brutal integration tax: perception models, sensors, controls, simulation, compute, and robot hardware all have to work together. Bringing agents into that development loop could reduce the engineering time required to configure and adapt robotic systems even before the robots themselves become materially more autonomous.

What’s next
The key question is whether agent-generated robotics work survives the harder standard of hardware-in-the-loop testing and production safety. NVIDIA is already positioning Isaac ROS alongside simulation and real world manufacturing deployments, making deployment quality and not just coding speed the next benchmark.

What happened
Alphabet owned Intrinsic open-sourced Intrinsic Core, a ROS compatible toolkit drawn from software it uses in manufacturing deployments. The release includes a hardware agnostic real-time control framework, digital twin, motion and grasp planning, simulation services, and an open machine tending reference design for industrial automation.

Why it matters
Industrial robotics teams repeatedly rebuild the same integration plumbing before they can work on higher value intelligence. Making reusable control, simulation, perception, and planning components open source could move more engineering effort toward adaptive AI applications and make physical AI deployments easier to reproduce across different robots and factories.

What’s next
Intrinsic is starting with manipulation-heavy industrial use cases such as CNC machine tending, not promising a universal robotics stack. Adoption by integrators, manufacturers, and the broader ROS ecosystem will determine whether Core becomes shared infrastructure or simply another useful toolkit.

Today’s through-line is systems economics: frontier intelligence is getting cheaper, agents are gaining purpose built execution and governance layers, and robotics infrastructure is becoming more modular and open. The competitive advantage is shifting from simply having the smartest model toward making autonomy affordable, permissioned, persistent, and deployable across software and the physical world.

Stress-Test the Agent Stack

Pick one real task and run it through two setups from today’s issue:

Then compare: cost, interventions, failures, and completion quality.

The point isn’t which model wins. It’s whether runtime + routing + permissions matter more than raw model intelligence.