On 1 October 2026 Google launched Guided Vision in Gemini Live on Android. It gives blind and low-vision users real-time conversational help with what their phone camera sees: live audio descriptions, spoken cues to reframe the camera, reading fine print, finding objects and describing colours and patterns. Google built it with accessibility communities and Aira, a visual interpretation service, drawing on more than 1,000 Aira Trusted Testers and what it calls tens of thousands of hours of training data.
That is a genuine accessibility gain, and it deserves to be said plainly before any comparison. The question for an Australian business is narrower: does a feature like this change which model partner you should build your own customer-facing tools on? Our view is that it mostly does not, and that the more useful lesson is in how Google built it.
Is Claude or Gemini better for accessibility?
They answer different questions. Guided Vision is a finished consumer feature inside Google's Android app, used directly by a person with low vision. When a business builds its own inclusive service, it is choosing a model to sit inside its own product, and the deciding factors are how that product is designed, tested with disabled users and governed. On that question, Claude and Gemini should both be tested on your actual tasks.
In other words, a person choosing an assistive app and a business choosing a model for its own service are making two separate decisions. Mixing them leads to vendor choices made for the wrong reason.
What Guided Vision gets right
Reading Google's announcement as a design case study is more useful than reading it as a scorecard. Several choices stand out and transfer to any build, whichever model sits underneath:
It was built with the people who will use it, through Aira and its tester community, rather than tested on them after launch.
It plugs into tools users already rely on, TalkBack and Accessibility Shortcuts, instead of asking them to learn a new interface.
It runs on Android 9 and later, so it reaches older and cheaper phones, not just new flagships.
It states its limits: Google says it is not a medical device, mobility aid or navigation tool, and that it can make mistakes.
That last point matters most for businesses. A clear statement of what a tool should not be relied on for is part of a responsible accessibility product, not fine print.
What Australian teams building inclusive tools should evaluate
If you are building a customer-facing service with vision or voice features, for example a claims app that reads a photographed document aloud or a support assistant that describes a product label, the model is one part of a longer checklist.
| Area | Question to ask | Why it matters |
|---|---|---|
| Co-design | Have disabled users shaped the design from the start? | Late testing finds problems too late to fix cheaply |
| Existing assistive tech | Does it work with screen readers and system shortcuts? | Users should not have to learn a new interface |
| Error handling | What happens when the model is wrong or unsure? | A confident wrong description can cause harm |
| Stated limits | Do users know what not to rely on it for? | Sets expectations and reduces risk |
| Privacy | Where do camera images and audio go? | Images can carry personal information under the Privacy Act |
| Model choice | Does the model handle your real images well? | Test with your own content, not demos |
Claude can interpret images, and we have written about how it handles inconsistent invoices, forms and IDs. That does not make it automatically the right model for every accessibility feature, and the same is true of Gemini. Test both on a sample of the images and requests your customers will actually send.
Why one platform feature is not a reason to lock in
Guided Vision lives inside Gemini Live on Android. Your customers use iPhones, Android phones, desktop browsers and phone calls. If your accessibility approach depends on one vendor's consumer app, you inherit that vendor's platform limits and roadmap. An accessible service has to work across all the ways your customers reach you.
A more durable approach keeps the accessible experience in your own product and treats the model as a replaceable component. Then your choice of Claude, Gemini or something else can be revisited on evidence, without redesigning the service each time a vendor ships a new feature. We made a similar point about Gemini Live's avatar: the interface a vendor adds is not the same as the decision about which model runs your service.
Where accessibility work usually starts
For most mid-market firms, the first accessibility win is less dramatic than live camera description. It is making the existing website, forms and documents work with screen readers. Our guide to running WCAG audits with Claude Code covers that groundwork. An illustrative budget for a mid-sized Australian site is $15,000 to $35,000 for an audit and fixes, before any new AI feature is considered. That figure is illustrative and depends heavily on the size and age of the site.
Get that base right, then add AI features where they remove a real barrier for real users, designed with those users.
What not to conclude
Do not conclude that Google now owns accessibility, or that Claude falls short because it has no matching consumer feature. Equally, do not treat any AI description tool as a safety device. Google itself is clear that Guided Vision can make mistakes. For a business, the standard is the same as for any customer service: tested with the people it serves, honest about its limits, and supported by a human path when it fails.
If you are planning an inclusive customer-facing tool and want help choosing and testing the model behind it, see our services or contact us.



