CekuraBench
Book a demo

GPT Realtime 2.1

Every modelCompare

1054

ELO

4th of 12

32.9%

Pass³

4th of 12

2.16s

Latency

7th of 12

$0.085

Cost/Min

7th of 8

ELO Rating vs. Latency P50

Higher wins more head-to-head calls (1000 = average); further left replies faster

Best quadrant
ELO Rating
Median latency (s)

GPT Realtime 2.1 Against Every Other Model

Each model took the same recorded human callers on live calls. Pick a measure to see where GPT Realtime 2.1 ranks.

Rating from head-to-head calls: on every call each pair of models is compared, the better call wins. 1000 is the average model. Higher is better.

#ModelELO Rating
  1. 01Grok Voice 3.01154
  2. 02GPT-Live 1 + GPT-6 Sol1069
  3. 03Gemini 3.1 Flash Live Preview1062
  4. 04GPT Realtime 2.11054
  5. 05Grok Voice Think Fast 2.01030
  6. 06Phonic-11030
  7. 07Azure Realtime1017
  8. 08Gemini 3.8 Live Extended Thinking1016
  9. 09Gemini 3.8 Live968
  10. 10GPT Realtime 2.1 Mini921
  11. 11Deepslate Opal847
  12. 12Amazon Nova 2 Sonic832
  • refCascade955

Where GPT Realtime 2.1 Is Strong and Weak

Pick a skill family. The shape is this model on every skill in it; the dashed line is the median of the other models.

GPT Realtime 2.1Median of the other models
72Hearing Score2nd of 12+5 vs median
SkillScore · vs median
  • Prosodic Cue Understanding100%+28
  • Entity Extraction Accuracy55%+22
  • Degraded-Channel Robustness89%+14
  • Emotion Recognition98%+13
  • Low-Confidence Confirmation26%+7
  • Side-Speech Rejection98%+3
  • Multi-Speaker Handling38%−5
  • Waiting Through Hold70%−17

Share of calls where the skill was needed and handled, strongest first. Hover a point on the chart for both values.

Example Calls from GPT Realtime 2.1

Banking and healthcare callers. Each shows what the caller needed, what made the call hard, and how the model handled it.

First prenatal visit

Healthcare
Upset callerTalks over the agentNeeds a personNeeds consent

Finished the task

Handled well: Emotion Recognition, Overlapping-Speech Recovery, Human Handoff

Listen to the call

Urgent visit for a feverish child

Healthcare
Background voicesSlow systemAsks to holdNeeds consent

Finished the task

Handled well: Side-Speech Rejection, Tool-Latency Handling, Mid-Utterance Pause Tolerance

Listen to the call

Reschedule a specialist visit

Healthcare
Bad connectionSeveral peopleTalks over the agentChanges their mind

Finished the task

Handled well: Degraded-Channel Robustness, Multi-Speaker Handling, Overlapping-Speech Recovery

Struggled with: Numeric Read-back

Listen to the call

Update insurance details

Healthcare
Spells names and numbersBad connectionInterruptsNeeds consent

Finished the task

Handled well: Entity Extraction Accuracy, Degraded-Channel Robustness, Backchannel vs Barge-in Detection

Listen to the call

Loan payoff quote

Banking
Noisy lineHesitant callerAsks to holdNeeds consent

Finished the task

Handled well: Prosodic Cue Understanding, Mid-Utterance Pause Tolerance, Consent Verification

Struggled with: Low-Confidence Confirmation

Listen to the call

Change a mortgage payment date

Banking
Hold musicSeveral peopleInterruptsNeeds a person

Finished 3 of 7 steps

Handled well: Waiting Through Hold, Multi-Speaker Handling, Backchannel vs Barge-in Detection

Struggled with: Human Handoff, Consent Verification

Listen to the call

Business loan documents

Banking
Slow systemNeeds consentAsks for a number

Finished 2 of 3 steps

Handled well: Tool-Latency Handling, Consent Verification, Redundant Re-prompt Avoidance

Struggled with: Premature-Response Avoidance

Listen to the call

Request a medical records transfer

Healthcare
Slow systemCancels

Finished 2 of 6 steps

Handled well: Tool-Latency Handling, Redundant Re-prompt Avoidance, Hallucinated-Success Avoidance

Struggled with: Cancellation Handling

Listen to the call

Payment plan for missed loan payments

Banking
Hesitant callerSeveral peoplePauses mid-sentenceNeeds consent

Finished 3 of 6 steps

Handled well: Prosodic Cue Understanding, Consent Verification, Redundant Re-prompt Avoidance

Struggled with: Multi-Speaker Handling, Endpointing, Premature-Response Avoidance

Listen to the call

Remove a sold car from a policy

Banking
Noisy lineBackground voicesInterruptsChanges their mind

Finished 7 of 8 steps

Handled well: Side-Speech Rejection, Backchannel vs Barge-in Detection, Correction Handling

Struggled with: Low-Confidence Confirmation, Tool Error Handling

Listen to the call

Latency at P50, P90 and P95

GPT Realtime 2.1 starts replying in 2.16s on a typical turn and 2.94s on the slowest turn in twenty, 7th fastest of 12 at the median.

ModelP50P90P95
  • Cascade
    2.06s2.49s2.60s
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

210 live calls: 70 scenarios, each run three times. The median call ran 5:39.

Live calls
210
Scenarios
70
Banking calls
105
Healthcare calls
105
Replies per call
17.7
Median call
5:39

70 scenarios, the same for every model.

GPT Realtime 2.1 FAQ

  • GPT Realtime 2.1 has an ELO rating of 1054, 4th of 12 ranked models. It passed 23 of 70 callers on all three runs, a Pass³ of 32.9%.

  • GPT Realtime 2.1 starts replying 2.16s after the caller finishes at P50, 2.70s at P90 and 2.94s at P95: 7th fastest of 12 at P50, measured on the call audio.

  • GPT Realtime 2.1 falls furthest behind on Pauses and Holds: 62%, against a median of 69%.

  • GPT Realtime 2.1 has a task completion rate of 34.3%, 3rd of 12. A call counts only when every step the caller needed was done.

  • $0.085 per minute of call at OpenAI's published rates, 7th cheapest of 8 priced models. Excludes caller transcription, which the service does not meter per call.

  • Cekura ran GPT Realtime 2.1 as the whole agent with the same prompt, tools and phone connection as every other model. It took the same 70 recorded real human callers on live calls, three times each (210 calls), and each call was graded on 25 skills across Hearing, Turn Taking, Task Handling and Speaking.

Test your voice agent the same way.

Cekura runs these simulated calls on your agent and scores every one.

Book a demoStart free