CekuraBench

GPT Realtime 2.1

Every modelCompare

79.3%

Reliability

2nd of 12

1.94s

Median reply

6th of 12

$0.085

Cost / min

7th of 9

Reliability vs. response time

Higher and further left is better

Best quadrant
Reliability, pass³ (%)
Median response time (s)

For the wider field, see our pick of the best voice AI APIs for real-time audio.

GPT Realtime 2.1 against every other model

Scenarios that passed on all 3 runs, pass³. Higher is better.

#ModelReliability
  1. 01Grok Voice 3.081.7%
  2. 02GPT Realtime 2.179.3%
  3. 03GPT-Live 1 + GPT-6 Sol75.6%
  4. 04Azure Realtime70.7%
  5. 05Phonic v168.3%
  6. 06Grok Voice Think Fast 2.067.1%
  7. 07Gemini 3.1 Flash Live Preview62.2%
  8. 08Deepslate Opal57.3%
  9. 09Gemini 3.8 Live Extended Thinking52.4%
  10. 10Amazon Nova 2 Sonic50.0%
  11. 11GPT Realtime 2.1 Mini50.0%
  12. 12Gemini 3.8 Live41.5%
  • refCascade82.9%

Where GPT Realtime 2.1 struggled

GPT Realtime 2.1 falls furthest behind on background noise: 76% passed, against a median of 81%.

Below the medianPassedMedian
  • Background noise
    76%median 81%
Mistakes in saved data
  • Wrong value10 · 34%
  • Field left out9 · 31%
  • Extra call6 · 21%

Fields it got wrong most: route_id, phone, destination

Response time at P50, P90 and P95

GPT Realtime 2.1 starts replying in 1.94s on a typical turn and 2.64s on the slowest turn in twenty, 6th fastest of 12 at the median.

ModelP50P90P95
  • Cascade
    2.17s2.51s2.62s
Seconds from the caller finishing to the agent replying. The bar is the typical turn, P50.

GPT Realtime 2.1 head to head

Every comparisonAll 12 models

How much audio GPT Realtime 2.1 was tested on

246 live calls: 82 scenarios, each run three times. The median call ran 1:55.

Live calls
246
Scenarios
82
Booking calls
177
Medicare calls
69
Replies per call
8.6
Median call
1:55

82 scenarios, the same for every model. See every test case

GPT Realtime 2.1 FAQ

  • GPT Realtime 2.1 passed 65 of 82 scenarios on all three runs, 79.3%, 2nd of 12 realtime models. Across all 246 calls, 89.0% passed every check.

  • GPT Realtime 2.1 starts replying 1.94s after the caller finishes at the median, 2.46s at P90 and 2.64s at P95. That is 6th fastest of 12 at the median, measured on the call audio.

  • GPT Realtime 2.1 falls furthest behind the field on background noise: 76% of calls passed, against a median of 81%. When it saved data wrong, the most common mistake was wrong value, 34% of them.

  • GPT Realtime 2.1 saved everything exactly right on 92.0% of the calls that save data, 1st of 12. On Booking calls, 95.9%. On Medicare calls, 81.8%.

  • $0.085 a minute of call at OpenAI's published rates, 7th cheapest of 9 priced models. Excludes caller transcription, which the service does not meter per call.

  • Cekura ran GPT Realtime 2.1 as the whole agent in the same open-source Pipecat pipeline as every other model, with the same prompt, tools and connection. A simulated caller worked through 82 scenarios on live calls, three times each: 246 calls for GPT Realtime 2.1. A call passes only when the agent said the right things and saved the right data.