Amazon Launches Bee Frontier and Eclipse: Prices, Subscriptions and Privacy Explained

Amazon has introduced Bee Frontier and Bee Eclipse, expanding its personal AI range with a wrist-worn device and a magnetic phone attachment. The October 15 announcement combines new hardware, paid integrations and privacy promises that deserve careful examination alongside a published security assessment.

Amazon announced two new Bee products on Thursday, October 15: the $74.99 Frontier, available now, and the Eclipse, coming soon from $99.99. The company announcement also introduced connections to Alexa+ for spoken access to summaries and to Amazon.com for shopping recommendations based on information customers choose to share.

For readers catching up on October 16, the central development is a broader attempt to turn everyday conversations into useful digital records. The hardware captures context; the service organises it. Those are distinct parts of the experience, and understanding the difference helps explain both the subscription model and the questions surrounding sensitive information.

Frontier puts conversation capture on the wrist or clothing

The Frontier product page describes a device that can be worn in a wristband or attached to clothing with the included clip. It lists dual microphones, noise filtering, IP67 water resistance and USB-C charging. The company advertises up to seven days of battery life, a specification that remains a manufacturer claim rather than a result from testing by Smashology.

The controls are deliberately simple. A single press starts or stops capture, a long press creates a voice note, and a double press bookmarks a moment. A green indicator signals that the device is on. The product documentation also describes a configurable silence timeout, with options ranging from five minutes to twenty-four hours.

Those controls matter because a memory aid has to fit into conversations without becoming difficult to manage. In practical evaluation, the useful question is whether someone can tell when it is active, stop it reliably and retrieve the moment they intended to save. A long battery-life claim alone does not answer those usability questions.

Eclipse offers a different place for the microphones

Bee’s Eclipse specifications describe a credit-card-shaped, 4.6-millimetre-thick attachment with MagSafe compatibility. A metal ring accessory is included. The product is intended to process in-person conversations and both sides of phone calls when activated, with a white status light and a button to start or stop it.

Bee lists roughly three weeks of battery life and a USB-C port. Its published controls also include voice notes and bookmarks. The company has announced a price, but the launch material does not give a firm release date. Readers should distinguish an announced product from one already available to order.

The design creates a different everyday tradeoff from a wristband. A phone attachment puts the device with something many people already carry, whereas a wearable remains attached to the person. Which arrangement captures an intended conversation more effectively will depend on actual use. Product dimensions and renderings cannot establish performance in a crowded room or during a difficult call.

The subscription changes the cost of the complete experience

The Bee Pro plan comparison lists a free tier with conversation processing, summaries and fourteen days of history. Pro costs $11.99 per month or $99.99 annually, with a thirty-day trial for new customers. Its features include longer history, desktop access, conversation templates, suggested tasks and connections to services such as Gmail, Outlook and calendars.

The same page says people who purchased a Bee Pioneer before October 15, 2026 retain their existing feature set without charge. Newly introduced capabilities such as Bee for Desktop fall under Pro. That qualification is relevant when comparing what an existing owner keeps with what a new buyer receives.

A hardware-only price therefore describes only one purchasing decision. Someone mainly seeking short-term summaries may evaluate the free tier differently from someone relying on integrations and an ongoing searchable archive. Comparing the intended workflow with the plan table is more informative than assuming every advertised feature is included with the device.

Connections make the data boundary more important

Bee’s current service overview presents searchable conversation summaries, a timeline of memories, suggested next steps and optional connections to mail, calendars, photos and health information. It also describes an Amazon.com integration that uses insights to suggest shopping items, and Alexa access to Bee information. Several functions require Pro.

These connections are central to the idea of an assistant that can act on remembered context. They also change what needs to be evaluated. A summary that remains inside one service is a different workflow from a summary used to draft an email or inform a recommendation elsewhere.

That distinction is an analytical point about the product’s design, not evidence that a customer has shared anything by default. The relevant questions are which connection the customer enables, what information it receives and how that permission can be withdrawn. An integration’s usefulness and its information requirements should be assessed together.

Amazon says audio is discarded, while text remains protected

In its explanation of Bee Private Compute, Amazon says captured audio is transcribed in real time and discarded. It says transcripts and derived information are encrypted with keys created and controlled by the customer’s device. Before sharing information with a server, the device checks that the receiving environment runs approved software.

The company describes isolated processing designed to prevent Amazon and Bee from accessing customers’ conversation content outside the authorised process. It also says customers choose what to share with connected services and can stop sharing. These statements describe Amazon’s architecture and commitments; they should be attributed to the company rather than presented as independently proven guarantees.

Deleting audio and retaining a transcript are different things. A transcript can preserve much of a conversation’s meaning even when its original sound is gone. Consequently, evaluating this service requires attention to the treatment of text, summaries and extracted tasks as well as the handling of microphone input.

The technical paper describes verified processing in the cloud

The 2026 Bee Private Compute paper describes device-generated root secrets and limited processing keys released after the receiving runtime is verified. Data is encrypted before it leaves the device and decrypted inside attested confidential virtual machines. The architecture also describes short-lived keys, transparency mechanisms and networking restricted by service identity.

In plain language, attestation is intended to give the device evidence about the software environment receiving the data. The paper’s design goes beyond simply promising that operators will avoid reading it. It attempts to make access depend on a defined processing path. The authors explicitly say the publication reflects the architecture at its publication date and may be updated.

This is still cloud processing. A product page’s reference to on-device processing should not be interpreted as meaning that every AI operation happens entirely inside the wearable or phone. The technical documentation describes a system with both a customer device and remote computing components, each playing a different role.

The infrastructure is only one layer of the system

Amazon says Bee Private Compute builds on AWS Nitro and NitroTPM. The AWS description of Nitro explains that it separates traditional virtualisation functions across dedicated hardware and software. Those functions include work involving computing, storage and networking. AWS presents that separation as part of the infrastructure’s security design.

This background helps locate the privacy claim. Infrastructure can provide isolation mechanisms, while the application still needs to handle keys, permissions and information flows correctly. Readers should therefore examine the Bee-specific paper and assessment alongside the description of the underlying platform. A statement about one layer cannot by itself establish that every feature built above it has been implemented correctly.

A security review records both findings and fixes

Trail of Bits lists an eight-week review commissioned by Amazon in its public audit catalogue. The assessment with fix review, dated September 9, 2026, reports twenty-one findings. Its final summary says seventeen were resolved, two partially resolved and two unresolved after the reviewed fixes.

The remaining unresolved items were classified low severity and informational. Two medium-severity findings were partially resolved. The report includes Amazon’s explanations for some remaining decisions, including the scope of developer tokens and the treatment of third-party integration data. It would be inaccurate to describe all twenty-one original findings as still open.

The review also explains that checking fixes involves specific code and configuration areas rather than a comprehensive re-analysis. These are dated assessment results, not a measurement of every component running at the moment a reader opens this article. Publication of a review provides useful evidence and a way to scrutinise claims; it does not eliminate the need to understand its scope.

The practical test is the quality of the remembered information

Security and usefulness answer different questions. Even a carefully controlled processing path does not establish whether a summary captured the correct date, speaker or commitment. An evaluation should compare the output with the actual conversation and check whether uncertain details are represented faithfully.

For example, a tentative suggestion to meet next week should remain distinguishable from an agreed appointment. A draft follow-up should preserve that distinction before anyone sends it. These are illustrative tests for conversation assistants, not reports of an observed Bee error. Smashology has not conducted hands-on accuracy or battery testing of these new devices.

Other participants also matter. Amazon says users are prompted to obtain permission from call participants when recording begins. Clear communication about when a device is active remains part of making the experience understandable to the people around it, alongside the product’s visible controls.

The October 15 launch supplies concrete hardware and subscription details, plus documentation readers can examine. The next evidence to watch is real-world performance: whether the devices reliably preserve useful context, whether customers understand their sharing choices and whether the available controls work as expected in daily use.

Explore more Smashology Tech reporting and our coverage of Europe’s latest AI science initiative.