GPT Realtime 2 for restaurant phone ordering: per-feature scorecard against a real 102.36 second call

What we are grading against

Vendor benchmarks are measured on 22 or 24 kHz studio audio in a quiet room. Restaurant phone calls arrive over the public switched telephone network as 8 kHz mono encoded in G.711, with kitchen ambient on one side and Bluetooth car audio on the other. The MOS score on a vendor page does not survive any of this. So we keep one fixed reference call and run every model swap against it.

The reference call is public/audio/dennys-order.mp3, 102.36 seconds, 16-bit stereo. The customer is on the left channel, the AI is on the right. The first three captions are the agent saying “Hi.” then “This is Denny on a recorded line.” then “What can we get for you?” The customer answers at 5.36 seconds: “Hi. Yeah. Can I get one Lumberjack Slam and one Coke?” The agent then asks for the egg style and bread choice, the customer modifies, and the order commits. This is the call we benchmark every new voice model against.

What gpt-realtime-2 actually shipped on May 7, 2026

Three models came out together: gpt-realtime-2 (the speech-to-speech model), gpt-realtime-translate (live cross-lingual speech), and gpt-realtime-whisper (streaming transcription). The Realtime API also exited beta and reached general availability the same day. The headline numbers for gpt-realtime-2 specifically: a 128,000 token context window (up from 32K), text plus audio plus image input, configurable reasoning effort across five levels, parallel tool calls, two new exclusive voices (Cedar and Marin), and a Big Bench Audio score of 96.6% (up from 81.4% on gpt-realtime-1.5). Audio pricing is unchanged from the previous model: $32 per 1M audio input tokens with cached input at $0.40, and $64 per 1M audio output tokens.

Below, one section per capability. For each one: what the feature does, what changes for a real restaurant phone call, and a verdict.

The per-feature walkthrough

The scorecard, side by side

Same eight features, scored against the actual Lumberjack Slam reference call and the production POS commit path. The right-hand column is what we typically see when a vendor says “we use the latest OpenAI model” without doing the per-leg work.

The single biggest behavioral change on a real call

If you take one thing from this page, it should be this: the new audible-preamble behavior is what makes unattended phone ordering feel like a real phone call rather than a button on an IVR menu. Before, the model would either answer immediately (and look unsure on a multi-step order) or pause silently while a tool call ran (and the caller would assume the line dropped, repeat themselves, or hang up). With audible preambles the model can say “one moment, checking the kitchen” while it actually checks the kitchen, and the caller stays engaged.

On the reference call this matters at the modifier turn. The agent asks “For your Lumberjack Slam, how would you like your eggs cooked, and what kind of bread would you like?” and then the customer answers. With a single-tool-call model, what comes next is silence while the POS write happens. With parallel tool calls and an audible preamble, what comes next is “one moment, sending that to the kitchen now,” followed by the order recap, and the caller never wonders whether the line is alive. That single shift is more important to the average restaurant phone call than any of the headline reasoning improvements.

Per-turn effort selection: the architecture decision most write-ups miss

Most coverage of gpt-realtime-2 reports a single latency number (1.12 seconds at minimal, 2.33 seconds at high) and treats reasoning effort as a global setting. On a phone line that is wrong. A typical restaurant call has 6 to 14 turns, and most of them are mechanical: the customer says an item, the agent confirms it, the customer adds a modifier, the agent confirms, the customer says they are done, the agent reads back. Those turns want minimal effort because the answer is a tool call and an acknowledgement.

The expensive turns are the ones where the customer says something that has to be reconciled against the order so far: a half-and-half pizza modification mid-stream, a substitution on a combo, a clarification on what counts as “the dinner deal.” Those want medium. The right shape is per-turn effort selection: a small classifier (or the model itself, with a constrained tool) decides per turn whether this one is mechanical or ambiguous, and dials the effort up only where the reasoning will pay for itself in fewer round trips.

The cost of getting this wrong is concrete. xhigh effort on every turn turns a 90 second call into a 130 second call, and you can hear it. Minimal effort on every turn produces wrong half-and-half splits and bad combo modifications, and you can read those in the next morning's POS audit.

What does not change for the operator

Pricing. PieLine pricing is $350 per month for up to 1,000 calls, $0.50 per call after that, with a money-back guarantee for the first month. The underlying model SKU is our problem to absorb. Per-minute pricing from some competitors reprices on every OpenAI announcement and tends to spike exactly when the line is busiest. Holding the bill stable so the operator can budget against covers, not against a vendor's release schedule, is part of why we chose this pricing shape.

Onboarding. Same-day onboarding still applies. The AI builder still scrapes the menu, maps it to POS item IDs, builds the modifier ontology, and tunes the barge-in policy for kitchen ambient. The model swap from gpt-realtime to gpt-realtime-2 is invisible to the operator. They pick up the phone the same way the day before and the day after; only the call quality changes.

Operator workflow. The dashboard, the per-call audit shape (audio, captions, envelope, ontology, POS receipt), and the smart-transfer flow to a human staffer with full context are unchanged. The new model fits behind the existing harness instead of replacing it.

What an operator should ask any vendor right now

If you are evaluating a restaurant phone AI vendor in May 2026, the gpt-realtime-2 release gives you a sharp tool. Three questions are enough to separate the vendors who did the work from the ones who updated their landing page.

First: are you on gpt-realtime-2, and do you select reasoning effort per turn? “Yes and no” is fine if they explain why a single global level fits their book of business. “We always use the latest version” with no per-turn answer means they have not run the eval.

Second: how do you handle the silent gap during a POS write? If the answer is “we use parallel tool calls with audible preambles,” they have done the work. If the answer is “our model is fast enough that there is no gap,” ask them to play a real call recording on a multi-modifier order.

Third: does my monthly bill move when OpenAI ships a new SKU? If yes, you are buying volatility. The model SKUs will change at least four times a year for the foreseeable future. Your monthly bill should not.