Amazon is developing a more capable, agentic version of Alexa under the codename "Moonraker," designed to execute multi-step tasks across simultaneous requests, according to internal planning documents reviewed by Business Insider. The publication's July 2026 report, previously unreported before it broke, describes a system capable of chaining actions in a single interaction — for instance, booking a ride and texting a friend without separate invocations Business Insider.

The economics are the story's sharpest edge. Amazon's internal projections put GPU costs for Moonraker at more than $100 million in 2026 alone, a figure that has reportedly prompted some senior employees to argue the company has already overspent on the models running the current Alexa+ generation Engadget. Internal documents cited in the reporting suggest Amazon leadership is weighing whether to delay Moonraker outright or scale back its ambitions rather than absorb that compute bill on the current timeline.

A separate document dated to late 2025 sketches the technical shape of the effort in more granular terms: hundreds of NVIDIA GPUs earmarked for testing, paired with an Anthropic Sonnet model handling advanced reasoning and visual response generation. That detail matters for two reasons. It confirms Amazon's continued reliance on Anthropic — in which it has invested heavily — for frontier-model capability it either cannot or chooses not to replicate in-house. And it signals that Moonraker's ambitions extend beyond voice, into multimodal reasoning that can parse and respond to visual input, a capability current Alexa+ does not advertise.

Context here matters as much as the codename. Alexa+ launched nationwide in the US at the start of 2026 and remains in Early Access in the UK as of this writing Engadget. Amazon's own announcement of the product, published in February 2025, framed its defining feature as agentic autonomy — the ability to "navigate the internet in a self-directed way to complete tasks on behalf of users" Amazon. Moonraker, then, is not a pivot so much as an escalation of a capability Amazon had already committed to publicly well over a year before the Business Insider report surfaced.

That escalation is running into friction on the ground. Users have reported Alexa+ struggling with basic requests that earlier, non-generative Alexa versions handled reliably Engadget. This is the crux of the tension inside Amazon, as described in the reporting: a product still working out reliability on single-step commands is simultaneously being asked, in the form of Moonraker's planning documents, to handle concurrent multi-step tasks with materially higher inference costs per interaction.

The reliability gap is not unique to Amazon, and it is worth situating in a pattern that has recurred across the industry's shift from deterministic voice assistants — built on constrained intent-recognition and slot-filling — to generative, LLM-backed ones. Deterministic systems failed predictably and narrowly; generative systems fail less predictably but often more expensively, since a botched agentic chain (the wrong ride booked, the wrong contact texted) carries higher stakes than a misheard timer request. Amazon's own internal skepticism about spend on the current Alexa+ models suggests the company is aware it has not yet resolved that tradeoff before contemplating a more ambitious, more expensive successor.

Whether Moonraker ships as scoped, gets delayed, or is scaled down is presented in the sourcing as an open internal debate rather than a settled roadmap. That uncertainty is itself informative: it indicates Amazon does not yet have confidence that agentic reliability at scale justifies GPU spend north of $100 million for a single fiscal year, at least not on the timeline originally planned. Anthropic's role as the reasoning backbone for testing also raises a longer-term question the reporting does not resolve — whether a shipped Moonraker would run production traffic on Sonnet-class models at that same compute cost, or whether Amazon would look to its own Nova models to bring the economics down before any wide release.

For consumers, the near-term signal is more modest than the codename implies: no launch date, market, or device tier has been reported, and Alexa+'s own rollout — nationwide in the US, still Early Access in the UK — suggests Amazon is already managing a staggered, cautious expansion of agentic voice AI before adding a more compute-intensive layer on top. The Moonraker reporting reads less as an announcement than as a snapshot of a company negotiating, in real time, how much reliability an agentic assistant needs before its cost becomes defensible.