Microsoft rebuilds Copilot: always-on agent, metered per-use billing

Chat, coding and a persistent agent become one app as Copilot's advanced tools drop the flat fee. Plus LangChain search, Perplexity Photon, Anthropic's $11.6B.

Nowline SEP 26 1:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • One Copilot, three doors — and an agent that never clocks out

    Microsoft folds chat, natural-language app-building, and a persistent “Autopilot” agent (the renamed Scout) into a single app split into Home, Code, and Autopilot. Autopilot gets its own cloud workspace and identity, keeps running while you’re offline, and resumes work after days — you @mention it from Teams, Outlook, or a document.

  • The flat fee is over: Copilot's power features go metered

    Autopilot, Code, Cowork, and the advanced Astra and Fable models shift from subscription to usage-based billing — the standard Copilot license now covers only chat and Office integrations. Cowork runs $0.01 per Copilot Credit and an auto-router picks the model by cost, speed, and accuracy. Price it out before you wire agents into anything that loops.

  • Ship an app as easily as sharing a Word doc

    Copilot Managed Runtime hosts the apps the Code section generates inside your Microsoft 365 boundary, with Git version control, editable code, and an SDK for third-party tools. Managed Runtime and the SDK are in public preview now; Code reaches Frontier participants at the end of September, with M365 Premium/Pro subscribers later in 2026.

  • LangChain's Managed Deep Agents 0.8 gets first-party web search

    The production Deep Agents harness now ships with built-in web search from Parallel plus OAuth credential handling, so your agents can authenticate into tools and pull live results without you bolting on a separate search stack. It lands in the managed LangSmith tier.

  • Perplexity's Photon search API: ~230ms results, ~68% cheaper for agents

    Perplexity's new Rust-based “Fast Search” (Photon) returns results in roughly 230ms and cuts agent search costs by about 68% versus its prior API — a low-latency drop-in for RAG and agent loops where every retrieval hop adds up.

  • Elsewhere: Anthropic commits $11.6B to Akamai for compute

    Anthropic signed a seven-year, $11.6B cloud deal with Akamai (plus a roughly $9B option), handing over warrants for up to a 5% stake. It's one slice of a reported ~$517B in compute commitments this year — the capacity sitting behind the Claude API you build on, and the sustainability questions trailing it.