Voice agent on OpenAI's GPT-Realtime-2

What’s changing

OpenAI shipped a new generation of voice models on May 8, 2026, led by GPT-Realtime-2 — their first voice model with GPT-5-class reasoning. We’re upgrading Routiq’s voice agent to run on it.

Why this matters for clinic calls

Voice calls into a busy clinic are messier than text. Patients trail off mid-sentence, change their minds, talk over the receptionist, mix two requests into one breath. The current generation of voice AI handles the easy 80% well. The hard 20% — “actually, can I bring my son to that one too, and could I switch from Tuesday to Thursday?” — is where models still trip.

GPT-Realtime-2 is built to handle that hard 20% better:

• GPT-5-class reasoning means the model can think more carefully about multi-part requests instead of pattern-matching the first thing it heard.
• Parallel tool calls mean the agent can check availability AND look up the patient AND verify the appointment type at the same time — cutting the awkward “let me check that for you” pauses.
• 128k context window means the agent remembers everything said in a long call. The patient who said “I’m allergic to lavender” at minute 2 won’t be forgotten at minute 8.
• Better interruption handling at the model layer, compounding with the interruption fix we shipped recently.
• Adjustable reasoning effort (normal, high, xhigh) — we can tune how much the model “thinks” per turn based on whether speed or accuracy matters more for that flow.

OpenAI reports a 15.2% improvement on the Big Bench Audio benchmark over the previous generation.

What you’ll notice

In day-to-day clinic operations: fewer of those small, annoying moments where the AI misses something on a phone call. More natural pauses. Faster availability lookups. Better follow-through on complex requests. Same voice, smarter brain.

What we’re working through

The model is in OpenAI’s Realtime API now, so the access is unblocked. The work for us is:

1. Side-by-side eval against the current voice agent on real call recordings (with permission) — does GPT-Realtime-2 actually outperform on Routiq’s specific workload, or just on generic benchmarks?
2. Re-tuning the system prompts and tool descriptions — newer models often reward different prompting patterns.
3. Cost modelling — the new model is more expensive per token, but parallel tool calls and better first-pass accuracy may net out to fewer tokens per call.
4. Gradual rollout: pilot on a couple of clinics with high call volume, watch the eval-harness scores, then broaden.

If the eval numbers don’t beat what we have, we stay on the current model. The point isn’t to chase shiny new releases — it’s to ship the call experience that actually books more patients.