Assistants are leaving the chat box: Copilot drives desktop apps while model makers fill out price tiers.

The week in AI assistants

Two things moved at once. At the product layer, assistants stopped living only in a text box: Copilot now operates desktop applications, and agent managers are appearing in clients. At the model layer, the main vendors filled out price tiers rather than pushing a single ceiling.

The tooling around them, runtimes and routers, spent the week keeping up with each new model.

Leaders

GitHub Copilot added computer use for desktop apps, made agentic autofix consult stored Copilot Memory before proposing fixes for security alerts, and added Claude Opus, local sandboxing and chat updates in its September 21 release. The center of gravity has moved from the editor to the CLI and the standalone Copilot app.

Claude shipped Sonnet 5.5 as the faster, cheaper second model of the 5.5 family, a week after Opus 5.5, which was described in the feed as matching the prior frontier model at 40 percent lower cost. It also opened a public developer plugin portal with review and analytics, which widens the product beyond models into a managed directory.

Ollama exposed each model's thinking levels and defaults through its API, so applications can set reasoning effort per model, and added explicit model capabilities in model creation. Most of its release candidates harden the MLX runner.

Baseten added web search as a server-side hosted tool for Grounded Inference, reached 1.0 on its CLI and deprecated several model APIs. Retiring models while adding tools is a routine sign of a hosting layer maturing.

Wildcards

InvokeAI released 6.14 with Wan 2.2 video generation and multi-GPU support, then an alpha of version 7 that replaces the tabbed interface with a widget workbench and autosaving projects. The feed notes the new database is one-way, a notable cost for a local-first tool.

LobeChat canary builds are about agents more than chat: goal acceptance review, auto-named worktrees and task runs. It reads like an agent work manager in a chat client's clothes.

Themes that compounded

• New frontier models were absorbed by routers within days: opencode added a GPT-6 system prompt and Grok 4.7 support.
• Agents gained memory or goals at GitHub Copilot and LobeChat.
• Capability metadata moved into APIs at Ollama, which now exposes thinking levels.
• Model retirement kept pace with launches at Baseten.

Watch this week

Watch how GitHub Copilot ships after computer use, since its cadence is near daily, and whether the plugin portal at Claude shows early submissions. Watch Ollama for the 0.35.1 release candidates becoming stable, and InvokeAI for follow-up alphas on version 7.