Alexa debuted with the Amazon Echo speaker on November 6, 2014, and has passed through three stages since: a command-matching assistant, a platform of third-party skills, and a generative version, Alexa+, announced on February 26, 2025. By October 2026 Alexa+ is open to all customers in the United States and Canada. Amazon's science team describes the original design as four linked AI tasks: wake word detection on the device, speech recognition in the cloud, language understanding, and speech synthesis.
Key takeaways
- The Echo launched on November 6, 2014. Third-party skills arrived in 2015, and Amazon marks the Alexa Skills Kit anniversary in June.
- Only the wake word is detected on the device. After a possible match, a short audio snippet goes to a larger cloud model for confirmation.
- By 2020 Amazon counted more than 100,000 skills, which created a discovery problem that name-free skill selection was built to ease.
- Alexa+ replaced fixed command matching with large language models, according to Amazon. It reached all U.S. customers in February 2026 and all Canadians on June 11, 2026, free with Prime. Accuracy figures for it have not been published.
- Most performance numbers in this article come from Amazon itself and have not been audited independently.
The four-part design behind the first Echo
In a November 2019 retrospective marking the Echo's fifth anniversary, Amazon Science described the assistant as built on four AI tasks. A wake word detector runs on the device and listens for "Alexa." Automatic speech recognition then turns audio streamed to Amazon's cloud into words. Natural-language understanding works out what the words mean so the system can act. Text-to-speech turns the reply into spoken audio.
The split affects privacy and accuracy. The wake word model that runs locally has a small memory budget. In a November 6, 2020 post on Interspeech research, Amazon Science explained that a possible detection triggers a short audio clip to be sent to a larger cloud model, which confirms or rejects it. The clip includes half a second of audio from before the device's estimated start, so the cloud model hears the whole word even when the device's timing is slightly off. Amazon reported that a model fed device context, such as whether music was playing, improved the false-reject rate by 14.6 percent on average over a baseline. That is the company's own result, measured on its own data.
Skills: from a closed device to a platform
The first Echo shipped with Amazon's own features. In 2015, the first third-party skills appeared, and customers had to say a skill's name to use it. Amazon's developer blog marked the five-year anniversary of the Alexa Skills Kit, the Alexa Voice Service, and the Alexa Fund in June 2020, which places all three in June 2015. By July 2020 the company reported more than 100,000 skills, more than 700,000 developers, and more than 200 Alexa built-in products from other manufacturers.
A catalog that large brought a problem Amazon Science named directly: with tens of thousands of skills and names that were hard to remember, customers could not find what they needed. The company's answer was name-free skill selection, where the assistant picks a skill from the request itself. The same retrospective lists self-learning, in which Alexa corrects recognition and understanding errors from signals such as a customer rephrasing a request, without human annotators.
Milestones with dates
The table lists events that Amazon's own pages date or that a research paper documents. It is not a complete history.
| Date | Event | Source type |
|---|---|---|
| November 6, 2014 | Amazon Echo launches | Amazon Science |
| 2015 (June anniversary) | Alexa Skills Kit, Alexa Voice Service, and Alexa Fund begin; first third-party skills | Amazon developer blog |
| September 29, 2016 | Alexa Prize announced: USD 500,000 for the top team and USD 1 million for the university if a socialbot holds a coherent 20-minute conversation | Amazon press release |
| 2018 | Alexa Presentation Language launches, along with Alexa Blueprints for building skills without code | Amazon developer blog |
| 2019 | Alexa Guard launches; Alexa Conversations enters testing with a small group of developers | Amazon Science |
| February 26, 2025 | Alexa+ announced, built on large language models available through Amazon Bedrock | Amazon |
| November 2025 | Alexa+ early access begins in Canada | Amazon Canada |
| February 2026 | Alexa+ available to all U.S. customers; free for Prime members, USD 19.99 per month otherwise | Amazon |
| June 11, 2026 | Alexa+ available to all Canadians in English and French; free for Prime members, CAD 27.99 per month plus tax otherwise | Amazon Canada |
The Alexa Prize and open conversation
Command matching works when a request fits a known pattern. Open chat does not. In 2016 Amazon launched a university competition to push research on that gap. A 2018 paper by Amazon researchers and the participating teams, Conversational AI: The Science Behind the Alexa Prize, reports that sixteen teams built socialbots and received real user conversations and ratings to improve their systems. The teams worked on language understanding, context modeling, dialog management, and response generation. The paper is a description of the contest and its methods by the organizer, not an independent evaluation.
What five years of progress meant, in Amazon's numbers
The 2019 retrospective reports that wake word and speech recognition errors fell about fourfold since launch, and understanding errors about threefold. It adds that Alexa worked in 15 language variants in more than 80 countries and that more than 85,000 smart home products from over 9,500 brands worked with it. These are company-reported figures, and the article does not state the test sets or the way errors were counted. The same piece lists open problems in plain terms: skill discovery, resolving references such as "near there," coordinating several skills for one task, using models offline, and fairness and bias.
From skills to a language model
Alexa+, announced on February 26, 2025, replaces the middle of that pipeline. Amazon says it is built on large language models, with Amazon Bedrock named at the announcement, and is meant to take longer, looser requests and carry out tasks across services. Amazon's 2026 U.S. page adds that the architecture uses models from Amazon Nova and from Anthropic. No published error rates for Alexa+ were found, so its accuracy against the earlier skill-based assistant cannot be stated from public sources.
What changed since the 2025 announcement
The 2025 announcement described an early-access product. By October 2026 the situation looks different, and the dates below come from Amazon's own pages. In the United States, Amazon's page says the free Prime tier started in February 2026. Alexa+ now covers all U.S. customers on Alexa-enabled devices, and the Alexa app, at no cost for Prime members and USD 19.99 per month for others. A free, usage-limited text chat is offered to non-Prime customers. Amazon says tens of millions of people joined the early-access program and that they talked to Alexa+ more than twice as much as to the previous assistant.
Canada followed. Early access began in November 2025, and a note dated June 11, 2026 on Amazon's Canadian launch page says Alexa+ is fully available in English and French, free with Prime and CAD 27.99 per month plus tax without it. Amazon Canada reports that Canadians use Alexa more than half a billion times a month and that Alexa+ usage per customer is twice that of the earlier version. As of Amazon's July 21, 2026 update, early access was also running in Brazil, the United Kingdom, Mexico, Italy, Spain, Germany, Austria, and France, and the company says feature availability varies by region. Amazon puts the installed base above 600 million Alexa devices. All of these figures are Amazon's own and have not been independently audited.
FAQ
When did Alexa first become available?
Amazon Science gives November 6, 2014 as the Echo launch date. Developers could build skills from 2015.
Does Alexa process everything on the device?
No. Wake word detection runs on the device, while speech recognition and language understanding happen in the cloud. Amazon Science has also described a larger cloud model that double-checks wake word detections.
What is a skill?
A skill is a voice app built by Amazon or an outside developer. Early skills had to be called by name, and later name-free selection let the assistant choose one from the request.
Is Alexa+ a different system from the original Alexa?
Amazon describes it as a next-generation assistant built on large language models. The original design matched requests to a fixed set of capabilities.
Is Alexa+ free in 2026?
For Prime members in the United States and Canada, Amazon says yes. Without Prime it costs USD 19.99 per month in the United States and CAD 27.99 per month plus tax in Canada, and a limited free text chat is available on the Alexa website and in the Alexa app.
Related reading
- How an AI voice assistant is built
- Conversational interfaces and AI
- AI through the years
- Where AI technology began
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Alexa at five: Looking back, looking forward, Amazon Science, November 6, 2019
- Amazon Alexa's new wake word research at Interspeech, Amazon Science, November 6, 2020
- 5 Years of Building for Voice with the Alexa Skills Kit, Alexa Voice Service, and Alexa Fund, Amazon Alexa developer blog, July 17, 2020
- Introducing the Alexa Prize: USD 2.5 Million to Advance Conversational Artificial Intelligence, Amazon, September 29, 2016
- Ram A, Prasad R, Khatri C, et al., Conversational AI: The Science Behind the Alexa Prize, arXiv:1801.03604, January 2018
- Introducing Alexa+, the next generation of Alexa, Amazon, February 26, 2025 (page updated July 21, 2026)
- Alexa+ now available to everyone in the US and free for Prime members, Amazon, 2026
- Alexa+ now fully available across Canada, free for Prime members, Amazon Canada, November 20, 2025 (updated June 11, 2026)