Claude Beats ChatGPT for Voice AI Agents

Follow me on LinkedIn - AI, GA4, BigQuery

The following is the virtual receptionist for a dental clinic, powered by GPT 5.1 (It's the one suggested by Retell AI).

0:00

/0:58

GPT-5.1 treats turn boundaries as a hard, prompt-driven property. It executes what you literally wrote. If a step in your prompt contains a multi-clause string ending in → end_call, GPT-5.1 reads that as one composite instruction: "say all of this, then end the call." Even with two explicit wait rules:

A voice agent that fires end_call prematurely loses the caller. There's no correction loop. The line is dead. The next time that customer needs your service, they remember the awkward AI experience and call your competitor.

The following is the same virtual receptionist (exact same prompt) for a dental clinic, powered by Claude 4.6 Sonnet (It's the one NOT suggested by Retell AI).

0:00

/1:00

Claude 4.6 Sonnet understand my instructions so well that even if I remove the following key rule, it still works fine:

Every question step requires a wait for the caller's answer before continuing

It means you can write shorter, cleaner prompts on Sonnet without losing reliability, which is one of the genuine advantages of using stronger models

You can engineer around it. It's just expensive in a different way.

To get GPT-5.1 to behave reliably like Claude, I've ended up adding things like:

Explicit "STOP. Wait for the caller's spoken response. Do not call any function until the caller has spoken." blocks on every question step.
Dedicated terminal steps for end_call, with Key Rules restricting which step can trigger it.
Branching logic written out explicitly because the model won't infer it.
Round after round of test calls because edits to one part of the prompt destabilise behaviour elsewhere.

That's real engineering cost.

Hours of authoring time per agent, plus ongoing maintenance brittleness. It doesn't appear in Retell's per-minute pricing, but it's not free.

Sonnet 4.6 typically needs none of this. The same flow runs in 30–40% fewer prompt tokens and works on the first deployment.

On Retell, GPT-5.1 runs at roughly $0.04 per minute. Claude Sonnet 4.6 runs at roughly $0.08 per minute. On paper, GPT wins on cost.

But in production, that's not always the case.

To get a cheaper model like GPT to behave as intended, you end up adding far more guardrails (which increases prompt size and, in turn, call cost) and far more testing. And there's still no guarantee the agent will behave as expected.

The 2x per-minute cost is real. The 95%+ reliability vs intermittent failure delta is bigger. A voice agent that doesn't understand the difference between a question and a closing tag isn't a cheaper voice agent. It's a broken one.

Don't take my word for it. Test it yourself.

The following is the same virtual receptionist (exact same prompt) for a dental clinic, powered by Claude 4.6 Sonnet (It's the one NOT suggested by Retell AI).

Claude 4.6 Sonnet understand my instructions so well that even if I remove the following key rule, it still works fine:

You can engineer around it. It's just expensive in a different way.

That's real engineering cost.

Related Articles:

Himanshu Sharma