Gemini Spark Opens Broader US Rollout with MCP-Driven Actions in Canva, OpenTable and Instacart
Spark crossed an important threshold this week: broader access has opened to US Google AI Ultra subscribers, and Spark can now execute tasks inside third-party apps via MCP, starting with Canva, OpenTable and Instacart. If you have been treating Spark as a demo, that calculus changes today. A 24/7 cloud agent that keeps running when your laptop is shut is now reachable for anyone paying $100 a month for Ultra in the US, with Workspace, Gmail and Docs already wired up.
The third-party action support makes Spark the first widely available consumer agent with MCP connectivity outside the developer surface, which matters if you have been deciding where to invest your automation effort. Expect Anthropic, OpenAI and xAI to respond on agent reach and pricing within weeks.
Codex Ships a Broad Update: Goals On by Default, Stronger Permissions, Plugin Discovery and a Better Windows Sandbox
If you use Codex inside the CLI, IDE extension or app, today's update materially changes how long-running work is handled. Goals are now on by default, backed by dedicated storage, and track progress across multiple turns, so you can hand Codex an objective and let it drive towards it for hours or even days without babysitting the chat history. Permission profiles gained list APIs, inheritance, managed requirements.toml support and a stronger Windows sandbox, which matters if you have been holding off rolling Codex out to a team because of access control concerns.
Plugin discovery is also easier to inspect, with marketplace-aware list output, installed versions and visible marketplace roots. Remote control now behaves more like a foreground command and reports machine status, making orchestrating Codex from CI or a parent agent meaningfully cleaner. The update was found on OpenAI's Codex changelog before mainstream coverage.
Claude Mythos Edges Closer to Public Release Through Claude Code and Claude Security
Anthropic's next-generation Mythos model is the story to watch this week. The model string claude-mythos-1-preview briefly surfaced in public Claude Code references, and a Mythos toggle was visible to some users before being pulled back. This matters because Anthropic publicly said in April that Mythos would not be released to the public in any form: the trail of UI strings, plus expanded Project Glasswing access, signals a reversal of that position.
The preview has already been used by around 50 partners to surface more than 10,000 high or severe-severity vulnerabilities, which gives you a sense of what to expect once the wider rollout lands. Watch the Claude Code release notes over the coming days if you want early access through that surface. This was found in Anthropic's own UI strings before mainstream press coverage caught up.
Windows 11 26H1 Preview Update Refreshes On-Device AI Components for Copilot+ PCs
If you run a Copilot+ PC or support users on one, today's non-security preview cumulative update KB5089570 for Windows 11 26H1 is worth noting. Microsoft has refreshed the on-device AI components covering Image Search, Content Extraction, Semantic Analysis and the Settings Model, all bumped to version 1.2604.515.0. These components underpin Recall, Click to Do and the natural-language Settings search, so the practical effect is faster on-device retrieval and better local-language understanding without round-trips to the cloud. It is a preview build, so test before broad deployment.
Industry Themes
The dominant pattern today is agents leaving the demo phase and starting to act on the user's behalf inside third-party software. Gemini Spark's broader US rollout with MCP-driven actions in Canva, OpenTable and Instacart, and OpenAI's Codex Goals shifting from experimental to default, both point the same way: persistent agents with durable objectives are becoming the assumed product shape rather than a research preview.
A parallel theme is the quiet repositioning of frontier security models for direct developer reach, with Anthropic's Mythos surfacing inside Claude Code rather than as a standalone launch. That trail was first visible through Anthropic's own UI strings before mainstream coverage caught up. For builders, the practical takeaway is that agent permission models, plugin discovery and on-device AI updates are now the surface area worth tracking, not just raw model benchmarks.