Leaked internal documents describe an upgrade that would let Alexa+ chain several related actions from one spoken request — book a ride while texting a friend — using an Anthropic Sonnet model.
2026-07-24 · by Moe Ameen
Amazon is working on a more agentic version of Alexa+ under the internal codename "Moonraker," according to a Business Insider report published in early July 2026 that cited internal planning documents. The upgrade is described as extending Alexa+ from single-task requests to compound, multi-step ones — understanding a request that bundles several related actions and carrying them out in sequence rather than stopping after the first. The example cited in the reporting is booking a ride while also texting a friend from one spoken instruction.
The documents reportedly describe testing an Anthropic Sonnet model for advanced reasoning and visual responses, and deploying "hundreds" of NVIDIA GPUs to run it. One planning document is said to have called Moonraker Alexa+'s highest-cost new initiative, projecting more than $100 million in GPU costs for 2026 alone. The reporting also noted internal tension — some Amazon leaders reportedly felt the company had already overspent on the models behind current Alexa, and the project's scope could be trimmed. Moonraker has not been announced publicly, and Amazon declined to comment on the leaked plans. Treat the specifics — the model, the GPU count, the cost figure, and the timeline — as unconfirmed internal targets, not shipped features.
This sits on top of a product that already leans agentic. Alexa+ is Amazon's generative-AI rebuild of the assistant, announced in February 2025 and rolled out to all US users in February 2026, powered by a combination of Amazon's own Nova models for fast voice responses and Anthropic's Claude for heavier reasoning. It can already take actions on your behalf — ordering food through partners like Grubhub and Uber Eats, booking rides, and navigating the web to complete errands — though users and reviewers have flagged reliability gaps, especially with third-party apps and remembering earlier context. Alexa+ is free for US Prime members and $19.99 per month for non-Prime users. Moonraker is the reported bet on making the "does it for you" part meaningfully more capable.
The interesting word in this story is "compound." Moonraker's whole pitch is turning one spoken request into several coordinated actions instead of one — book the ride and text the friend, from a single instruction. That is exactly the shape of the problem a creator faces, except the actions aren't errands, they're outputs: one idea should become a captioned short, an avatar cut, a carousel, an X thread, a blog post, and a newsletter, each reframed per platform and each scheduled. Alexa+ is building the agentic version of that for a person's daily tasks. Kompozy is the agentic version of it for a creator's publishing.
Concretely: point Kompozy at one source — a transcript, a rough note, a link — and it treats "make my content for this" as the compound request. It generates the full set in your Persona Brief voice, auto-captions the video, reframes each asset to the right aspect ratio, and queues everything to publish across the eight social platforms plus blog and email from a single review pipeline. Turn on autopilot and the compound request becomes a standing one: set your sources and cadence, and it generates, schedules, and publishes on a recurring basis without you assembling each post by hand. Moonraker wants Alexa+ to finish a chain of steps after you say one thing. Kompozy already finishes the content chain after you give it one input — and unlike a voice assistant boxed inside its own ecosystem, the finished work lands everywhere your audience actually is.
Moonraker is the internal codename, reported by Business Insider in early July 2026 from leaked planning documents, for a more agentic version of Alexa+. It would let the assistant handle compound, multi-step requests — carrying out several related actions from one spoken instruction, such as booking a ride while texting a friend — reportedly using an Anthropic Sonnet model. Amazon has not announced it publicly and declined to comment.
Alexa+ already takes actions on your behalf — ordering food through partners, booking rides, navigating the web — but largely one task at a time, and reviewers have noted reliability gaps. Moonraker is described as adding the ability to understand and complete compound requests that chain several related steps in sequence, plus stronger reasoning and visual responses.
No. It is an unannounced internal project as of this writing, and the figures reported — including more than $100 million in projected 2026 GPU costs — come from leaked documents, not a shipped product. Alexa+ itself is live in the US: free for Prime members and $19.99 per month for non-Prime users.
Moonraker generalizes the pattern of turning one instruction into many coordinated actions — for personal errands. Kompozy applies the same pattern to content: point it at one source and it generates video, carousels, threads, blogs, and newsletters in one brand voice via a Persona Brief, then schedules and publishes them across the eight social platforms plus blog and email, or runs the whole cadence on autopilot.