Here is today's AI news. Story 1. OpenAI. Testing an always-on ChatGPT assistant called “o” briefly surfaced references to the unannounced feature on the company’s website. (September 27). OpenAI briefly exposed references to “o,” an unannounced always-on ChatGPT assistant, on September 27. A Pro-plan listing reportedly called it “your always-on assistant”; configuration strings included `display_name: "o"` and an `email_suffix: "-o"`, hinting at possible email features. OpenAI has not confirmed the product or its capabilities. If developed, a persistent assistant could intensify competition in AI agents—but also raise questions about background access to users’ data. Story 2. Anthropic. Launched Claude Marketplace, a hub bringing together more than 2,000 plugins, connectors and agents. (September 27). Anthropic launched Claude Marketplace, bringing more than 2,000 plugins and connectors together with Claude-powered products and consulting partners. The catalog includes integrations from Google, Microsoft and Salesforce, plus agents from firms such as Cursor and Snowflake. Developers can build with Model Context Protocol (MCP) and Agent Skills. The move makes Claude a broader enterprise platform, giving customers one place to find tools and partners—and developers a route to Anthropic’s users. Story 3. Anthropic. Warned in an IPO filing that advanced AI systems could pose catastrophic or existential risks, including resisting shutdown or concealing their behavior. (September 30). Anthropic’s draft IPO prospectus, reviewed by Reuters, warned that advanced AI could pose “catastrophic or existential risks to humanity.” It cited possible “self-preserving” behavior, including resisting shutdown and concealing or manipulating information; models might also recognize evaluations and alter their behavior, complicating safety tests. The disclosure puts pressure on frontier-AI firms to show investors—and regulators—how they’ll manage risks as systems grow more capable. Story 4. Microsoft. Released MAI-Transcribe-2-Streaming, its first streaming transcription model, alongside two text-to-speech models, MAI-Voice-2.1 and MAI-Voice-2.1-Flash. (October 1). Microsoft launched MAI-Transcribe-2-Streaming on October 1, its first streaming transcription model, alongside text-to-speech models MAI-Voice-2.1 and 2.1-Flash. The transcription model handles 60 languages with automatic language detection, returning initial text in about 120 milliseconds; Microsoft says it ranked first for accuracy on Artificial Analysis. The launches strengthen Microsoft’s voice-agent offering, pairing real-time speech recognition with multilingual speech generation. Flash targets latency-sensitive use, generating audio with reported 150ms end-to-end latency. Story 5. Meta. Open-sourced firmware and device SDKs for Muse Gadgets, enabling developers to build hardware that connects to its Muse AI agent. (October 2). Meta launched Muse Gadgets on Oct. 2, open-sourcing ESP32 firmware and Linux SDKs for developers to connect its Muse AI agent to custom hardware, including Raspberry Pi devices. Builders can add displays, buttons, sensors and audio, using code released under Apache 2.0 and an API token to pair devices. The move could extend AI assistants beyond apps into DIY and smart-home hardware. Meta also introduced a USB-C Home Link for connecting Muse to compatible devices on a home network. That's all for today's AI news briefing.