Amazon Project Moonraker: What Developers Need to Know

Amazon is developing Project Moonraker, an upgrade to Alexa that would enable multi-step task execution from a single command, according to leaked internal documents. The project, which relies on Anthropic's Sonnet model for advanced reasoning, faces internal debate due to projected GPU costs exceeding $100 million in 2026. Amazon has not confirmed the project but has already released three AI-native SDKs for Alexa+ that developers can use now.

Amazon is building Moonraker. Leaked internal planning documents surfaced this week showing a project designed to turn Alexa into a true agentic assistant — one that handles compound, multi-step commands in a single request. The documents show GPU cost projections exceeding $100 million in 2026 and a test stack running on Anthropic’s Sonnet model. Amazon has not confirmed any of this. But the SDKs developers need to build for this future already exist and are open for business. What Moonraker Actually Changes The current Alexa+ handles one action per command. You ask it to book a restaurant, it books the restaurant. You ask it to text your friend, it sends the text. What Moonraker is designed to do is chain those together: “Book a table at Nobu and text Sarah that we’re on for 8pm.” One command. Two actions. Completed in sequence. That sounds trivial until you try building it. Multi-step task execution requires the model to parse compound intent, maintain context across tool calls, handle failure gracefully, and return a coherent result to the user — all without asking for clarification. This is the same problem Claude Code and OpenAI Agents are solving for developer workflows. Moonraker is Amazon’s bet on solving it for consumer voice. The test setup from the leaked documents leans on Anthropic’s Sonnet model for “advanced reasoning and visual responses” — which is telling. Amazon’s own Nova models power much of Alexa+, but Moonraker is apparently too demanding to run cheaply on first-party models today. The Three SDKs You Can Use Right Now Do not wait for Moonraker. Amazon already shipped three AI-native developer SDKs for Alexa+, and they are the foundation Moonraker will build on. Alexa AI Action SDK is the most mature. It connects your service to Alexa’s reasoning layer with real-time API calls, LLM-powered intent matching, and built-in personalization. If you have a bookable service — appointments, reservations, deliveries — this is your entry point. It is generally available. Alexa AI Web Action SDK does something genuinely useful: it lets Alexa navigate your existing website to complete tasks without you building a dedicated API. A user asks Alexa to find a plumber with weekend availability; the SDK handles the web interaction. It is currently in early access. Alexa AI Multi-Agent SDK is the most interesting for developers building specialized agents. It lets you run your own AI agent alongside Alexa — separate personality, separate interaction style — while the SDK handles coordination. Think of it as your agent getting distribution on 500 million Alexa-enabled devices. It is in preview; apply at developer.amazon.com/en-US/alexa/alexa-ai https://developer.amazon.com/en-US/alexa/alexa-ai . The $100M Problem Here is where you need to be careful. The same documents that reveal Moonraker’s capabilities also show internal debates about whether to build it at all. The projected GPU cost for Moonraker in 2026 exceeds $100 million, and Amazon’s own planning docs floated scaling back or delaying the project to manage those costs. Moonraker is not announced. It has no public roadmap. If you architect your product around capabilities that only Moonraker would unlock, you’re betting on an internal Amazon project that is still being debated. That’s a bad engineering decision. The smarter approach: build on the three SDKs that exist today. They support compound workflows through chained action calls already, just with more developer effort than Moonraker would require. Moonraker is additive upside, not the foundation. Where Alexa Sits in the Agentic Stack If you’re choosing where to build agentic experiences right now, the landscape looks like this: OpenAI Agents SDK leads on voice pipeline quality https://www.androidauthority.com/alexa-project-moonraker-project-3685647/ with native duplex audio and the best realtime mode. Claude Agent SDK wins on reasoning depth and computer use. Google ADK dominates enterprise GCP deployments. Alexa’s differentiation is distribution. Five hundred million Alexa-enabled devices. Consumer payment infrastructure through Amazon Pay. A decade of consumer trust for purchase and booking flows. No other agent platform has a hardware install base to match. The model may not be the best, but you don’t need the best model when your agent runs on the device already sitting on the kitchen counter. Build MCP-compatible tools regardless of which platform you prioritize. All four major agent frameworks now support the Model Context Protocol, which means one set of tool definitions can reach OpenAI Agents, Claude, Google ADK, and Alexa’s ecosystem with minimal rework. What to Do This Week If Alexa’s install base is relevant to your use case, review the Alexa AI Action SDK documentation https://developer.amazon.com/en-US/alexa/alexa-ai and map your service’s API surface to Alexa’s action model. Apply for early access to the Multi-Agent SDK now — preview slots are limited and Moonraker’s capabilities will land on top of this infrastructure when they ship. If your use case is developer tooling rather than consumer voice, Moonraker is not your story. Follow the Moonraker leak coverage https://thenextweb.com/news/amazon-moonraker-alexa-ai-agent-cost for timeline signals, but build for the platforms where agentic developer tooling is already mature.