GPT Realtime 2.1 is ahead on reliability, success rate, agent responses, data accuracy and response time, and GPT Realtime 2.1 Mini on stalled calls, interruption and cost.
Scenarios that passed on all three runs.
GPT Realtime 2.1, 29.3 points higher
Calls that passed every check.
GPT Realtime 2.1, 27.6 points higher
Calls where the agent's replies reached the expected outcome.
GPT Realtime 2.1, 5.7 points higher
Calls where everything the agent saved matched exactly.
GPT Realtime 2.1, 27.5 points higher
Calls where the agent went silent for 10 seconds or more.
Realtime 2.1 Mini, 0.8 points lower
How well it handled being talked over, 0 to 5.
Realtime 2.1 Mini, 0.01 higher
Median time to start replying after the caller finishes.
GPT Realtime 2.1, 0.15s sooner
One minute of call at the vendor's published rates.
Realtime 2.1 Mini, $0.065 cheaper
The small figure is the place among ranked models. The cascade is not ranked.
GPT Realtime 2.1 1.94s, Realtime 2.1 Mini 2.08s at the median.
GPT Realtime 2.1 passed 65 of 82 scenarios on all three runs, and GPT Realtime 2.1 Mini passed 41. GPT Realtime 2.1 is ahead, by 29.3 points.
GPT Realtime 2.1 saved everything exactly right on 92.0% of the calls that save data, and GPT Realtime 2.1 Mini on 64.5%. On Appointments, 95.9% against 88.8%. On Medicare, 81.8% against 0.0%. A call counts only when nothing is missing, wrong, extra or saved twice.
GPT Realtime 2.1 starts replying 0.15s sooner at the median. GPT Realtime 2.1 takes 1.94s at the median and 2.46s at p90; GPT Realtime 2.1 Mini takes 2.08s and 2.75s. Both are measured on the call audio, from the caller finishing to the agent starting to reply.
GPT Realtime 2.1 costs $0.085 a minute of call and GPT Realtime 2.1 Mini $0.020, at each vendor's published rates. GPT Realtime 2.1 Mini is $0.065 cheaper a minute.
Both ran as the whole agent in the same open-source Pipecat pipeline, with the same prompt, tools and phone line. A simulated caller worked through all 82 scenarios on live calls, three times each, 2,214 calls across the benchmark. A call passes only when the agent said the right things and saved the right data.
