As of July 2, 2026, the AI developments worth watching in the past 24 hours have focused on development tools and infrastructure: GitHub has clearly set the final deadline for shutting down Models, while further integrating open weighted models, visual understanding, and browser operations into Copilot; Anthropic and Meta have updated their agent collaboration and training storage bases, respectively.
GitHub Models will be completely discontinued on July 30
On July 1, GitHub announced in its official Changelog that GitHub Models will be fully retired on July 30, 2026. At that time, playgrounds, model directories, inference APIs, and BYOK endpoints will all go offline, including already active users. Brief interruptions for drills will also be scheduled on July 16 and 23. Projects still calling their inference API should immediately inventory dependencies and migrate, rather than just replacing the front-end entry point.
Kimi K2.7 Code entered the Copilot model selector
On the same day, GitHub announced that Kimi K2.7 Code officially entered Copilot, becoming the first open-weighted model in the model selector, hosted by GitHub on Microsoft Azure. It is gradually being opened to Pro, Pro+, and Max users, with Business and Enterprise following suit and closing by default, requiring administrators to enable the policy. For teams, adding new options does not mean direct switching; they still need to first verify price, code data governance, and output stability.
Copilot opens up both visual and browser tools
Copilot Vision is now officially available for all plans, allowing you to process images or PDFs in VS Code, GitHub web version, and CLI. VS Code's browser tool has also finished previewing; the agent can open the page, click, type, read console errors, and take screenshots. GitHub emphasizes that users' original tabs are not shared by default, and clear authorization is still required for camera, microphone, location, and clipboard reading; Businesses can also use domain whitelists and blacklists to restrict access.
Claude Code 2.1.198 enhances backend agent delivery
On July 1, Anthropic released Claude Code v2.1.198: Claude in Chrome became officially available. After the backend agent completes the code in the worktree, it can automatically submit, push, and create draft PRs; The new version also adds background task notifications and /dataviz skills, and fixes issues such as brief network outages and tasks getting stuck in Running. The focus of change is not on adding an order, but on long tasks beginning to have a more complete delivery loop.
Meta reworks AI training storage chain
Meta's engineering blog revealed a new BLOB storage architecture designed for large-scale AI training. Old systems' multi-layer metadata queries could accumulate to hundreds of milliseconds, slowing down GPUs during synchronous training; The new solution rebuilds the metadata layer, using memory, host flash, and regional flash to form layered caches, combining prefetching and on-demand loading to reduce cross-region transfers. It shows that the efficiency competition for frontier models has extended from GPU count to storage tail latency, data location, and research iteration times.
