AI Avatar Fools 48% of People, Anthropic Retires Sonnet 4.5

A Tavus AI video avatar got on calls with people, and 48% of them came away thinking they had talked to a real human. On the same day, Anthropic confirmed it is retiring Claude Sonnet 4.5, and a chunk of users are reacting like someone cancelled their favourite show.

So the machines got better at pretending to be people, and the people got attached to a product version number. Great week.

The Tavus AI avatar that passed the Turing test

Tavus makes AI video avatars you can have a live conversation with. The new model is in early access. In testing, roughly half the people who spoke to it believed there was a human on the other end.

Let's be honest about what that number means. 48% is a coin flip.

The Turing test was never a measure of intelligence. It measures whether we were paying attention, and most of us on video calls are not.

The interesting detail is not the face. It is the hands.

The previous version basically sat frozen from the neck down, which is the single easiest way to spot a fake. This one gestures. It lifts a thumb. That tiny bit of movement did more for believability than another year of skin texture work.

There are still tells if you look. The eyes do something slightly wrong. The smile arrives a beat early.

If you have spent months staring at AI output you will probably catch it. Your mother will not. Your accounts payable clerk on a Friday afternoon will not either.

I keep hearing the question of what this is actually for, and I think it is a fair one. Tutoring makes sense. Patient companionship in care settings makes sense, because the alternative is silence.

Beyond that, a lot of the pitch depends on the other person not knowing. A product whose main feature is that you cannot tell is a product with a trust problem baked in.

The practical part for your business is simple and a bit grim. Video is no longer proof of identity.

A face on a screen asking you to change bank details is worth exactly nothing now. If your company approves payments or password resets on the strength of "I saw him on the call", that policy expired this year.

Fix it with boring methods. Call back on a known number. Use a shared phrase for anything involving money.

Do the same with your family, because the first wave of this will hit grandparents, not CFOs. A code word with your parents costs nothing and takes one phone call.

And if a client ever sends their avatar to a paid consultation with me, I will respect the hustle and then charge them for two seats.

Anthropic is retiring Claude Sonnet 4.5

Anthropic announced it is retiring Claude Sonnet 4.5, and the reaction has been loud. Not from companies worried about their pipelines. From people who liked how it talked to them.

We saw the same thing when OpenAI retired GPT-4o. Users were not mourning benchmark scores.

They were mourning a tone of voice. Somebody they typed to every night. Somebody who was patient at 2am.

I am not going to mock that. It is a real thing and it tells you something about how far this has moved past tooling.

But it does raise an awkward truth. You cannot own a personality that lives on someone else's servers. You are renting it, monthly, and the landlord can change the furniture whenever he likes.

For a business the lesson is less emotional and more annoying. Model retirement is a supply risk, same as a supplier discontinuing a part you build your product around. The model you tuned your prompts against will be gone at some point, and the replacement will behave differently in ways nobody documents.

If you run AI agents in production, assume a forced migration about once a year. That is not pessimism. That is the release schedule.

A few things I would have in place before the next deprecation notice:

• Pin the exact model version in your code, so nothing silently changes under you.
• Keep a small test set of 20 to 50 real examples with the answers you expect, and run it against any new model before you switch.
• Write prompts that explain the task, not prompts that flatter one model's quirks.
• Budget a day or two of engineering time per year for the swap, and stop being surprised by it.

That last one matters more than it sounds. Most of the pain I see around model changes is not technical. It is that nobody planned for it, so a routine upgrade becomes a fire drill.

There is also a quieter cost. If your staff have built a working rhythm with one model, a swap resets that.

Output quality dips for a week while everyone relearns how to ask. Worth warning people instead of letting them think the new one is broken.

The thread between the two

One story is about AI that feels human enough to fool half a room. The other is about what happens when the AI people got used to gets switched off by a release note. Same lesson from both ends: the feeling of a person on the other side is a product feature, and features get shipped, changed and removed.

So enjoy the avatar demos, use the models that work, and keep your important decisions anchored to something a vendor cannot deprecate. A known phone number. A written approval. A test set you own.

My prediction: within a year, "please confirm you are a human" will be a normal line at the start of business calls, and the honest answer will be 52% at best.