This Week
One post per release. The same list is the newsletter, and there is an RSS feed if you would rather not give up an email address.
Releases
Roboflow Agent builds computer-vision apps
A vision assistant that picks models, designs workflows, and can take tasks from Claude Code, Codex, or Cursor over MCP.
Gemini 4 Argon is generally available
Google's frontier model for hard coding, science, and agent work is now in the API and for AI Pro and Ultra subscribers.
Meta releases Muse Wonder for agent work
Muse Wonder and Muse Wonder Full are Meta's new models for personal and professional agent tasks, available in the Meta Model API and Muse Code.
Claude Managed Agents can write their own workflows
A beta lets a managed agent split a large job into phases, run many agents, and combine the results in the background.
Google retires two Gemini Flash model ids
gemini-3.7-flash and gemini-3.5-flash now route to newer Flash models, and an old Deep Research agent shuts down on 23 October.
OpenAI adds Ultrafast mode for GPT-6.1 Sol
The Responses API can run gpt-6.1-sol with service_tier ultrafast so tokens arrive with less delay.
Claude Haiku 5.5 is Anthropic's new small model
Haiku 5.5 is the fastest, cheapest small Claude yet, with a 1 million token context window and adaptive thinking.
Intercom ships Fin 4 for support agents
Fin 4 updates the customer agent, the Operator, human review, and the Apex models used for support.
GPT-6 brings Intelligent UI to ChatGPT
ChatGPT can now mix text with charts, buttons, and small tools in one reply. Paid chats use GPT-6 Sol; Free and Go use GPT-6 Luna.