OpenAIAudio modelActive

GPT-Realtime 2.1

PopularityRank 256 of 875

41out of 100same as last month

Summary

GPT-Realtime 2.1 updates GPT-Realtime 2 for voice-agent workflows. OpenAI calls out better alphanumeric recognition, silence and noise handling, and interruption behavior. The model accepts text, images, and audio and returns text and audio. Configurable reasoning and tool use make it broader than a transcription-only model.

The API profile provides a 128,000-token context and up to 32,000 output tokens. OpenAI notes that higher reasoning effort can increase latency and output token usage. GPT-Live-Transcribe remains the separate transcription specialist, while GPT-Realtime 2.1 combines audio input and output with reasoning and tool calls.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 49 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Artificial Analysis Big Bench Audio speech reasoning (Show measurement, test conditions, and source)

Source value: 96
Score: 96
Metric: accuracy
Unit: %
Benchmark version: Big Bench Audio v1.2
Category: audio
Direction: Higher is better
Sample: 1000
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Native audio-to-audio on Big Bench Audio v1.2, 1,000 questions over four 250-item reasoning categories. Exact configuration: GPT-Realtime-2.1 High.

Source Artificial Analysis

Artificial Analysis Conversational Dynamics (Show measurement, test conditions, and source)

Source value: 95.7
Score: 95.7
Metric: conversational behavior score
Unit: %
Benchmark version: AA Full Duplex Bench v1 + v1.5 view
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Full Duplex Bench v1/v1.5 conversational behavior view (pause handling, turn taking, interruption and backchannel categories). Exact configuration: GPT-Realtime-2.1 High.

Source Artificial Analysis

Artificial Analysis τ-Voice task completion (Show measurement, test conditions, and source)

Source value: 45.7
Score: 45.7
Metric: Task Completion pass@1
Unit: %
Benchmark version: Artificial Analysis τ-Voice base task set
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Customer support agent task completion across airline (50), retail (114), telecom (114) scenarios, with tools and policy. Publisher's general protocol averages three trials; exact row GPT-Realtime-2.1 High uses 2 trials.
Context: The publisher specifically notes this exact configuration used two trials, not its general three-trial protocol.

Source Artificial Analysis

Speech Agent Arena task success rate (Show measurement, test conditions, and source)

Source value: 91.5
Score: 91.5
Metric: eligible-conversation task success rate
Unit: %
Benchmark version: Current Speech Agent Arena board retrieved 2026-10-15
Category: audio
Direction: Higher is better
Lower confidence bound: 87.7
Upper confidence bound: 94.3
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Correct final task-completing tool call on eligible conversations only; participant deviations and unverifiable cases are excluded. Exact configuration: GPT-Realtime-2.1 High.

Source Artificial Analysis

Artificial Analysis Big Bench Audio speech reasoning (Show measurement, test conditions, and source)

Source value: 87
Score: 87
Metric: accuracy
Unit: %
Benchmark version: Big Bench Audio v1.2
Category: audio
Direction: Higher is better
Sample: 1000
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Native audio-to-audio on Big Bench Audio v1.2, 1,000 questions over four 250-item reasoning categories. Exact configuration: GPT-Realtime-2.1 Minimal.

Source Artificial Analysis

Artificial Analysis Conversational Dynamics (Show measurement, test conditions, and source)

Source value: 92.7
Score: 92.7
Metric: conversational behavior score
Unit: %
Benchmark version: AA Full Duplex Bench v1 + v1.5 view
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Full Duplex Bench v1/v1.5 conversational behavior view (pause handling, turn taking, interruption and backchannel categories). Exact configuration: GPT-Realtime-2.1 Minimal.

Source Artificial Analysis

Artificial Analysis τ-Voice task completion (Show measurement, test conditions, and source)

Source value: 38
Score: 38
Metric: Task Completion pass@1
Unit: %
Benchmark version: Artificial Analysis τ-Voice base task set
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Customer support agent task completion across airline (50), retail (114), telecom (114) scenarios, with tools and policy. Publisher's general protocol averages three trials; exact row GPT-Realtime-2.1 Minimal uses 3 trials.

Source Artificial Analysis

Speech Agent Arena task success rate (Show measurement, test conditions, and source)

Source value: 89.4
Score: 89.4
Metric: eligible-conversation task success rate
Unit: %
Benchmark version: Current Speech Agent Arena board retrieved 2026-10-15
Category: audio
Direction: Higher is better
Lower confidence bound: 85.2
Upper confidence bound: 92.5
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-15
Methodology: Independent evaluation. Correct final task-completing tool call on eligible conversations only; participant deviations and unverifiable cases are excluded. Exact configuration: GPT-Realtime-2.1 Minimal.

Source Artificial Analysis

Cekura all-three-run scenario pass rate (Show measurement, test conditions, and source)

Source value: 79.3
Score: 79.3
Metric: share of scenarios passing all three runs
Unit: %
Benchmark version: Cekura Bench GPT Realtime 2.1 speech-to-speech page, published and modified 2026-10-16; result-run revision not stated
Category: audio
Direction: Higher is better
Sample: 82
Source type: independent-evaluator
Evaluator: Cekura Bench speech-to-speech benchmark
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. Cekura labels the evaluated model GPT Realtime 2.1. This is the complete phone-agent configuration in a Pipecat pipeline, with the same prompt, tools and connection across compared models, not a model-only score. Cekura reports 65 of 82 scenarios passed on all three runs (79.3%); each scenario was run three times, for 246 live calls total. The exact deployed agent commit and OpenAI API serving revision are not stated. The evaluator README describes scenario-stratified bootstrap intervals for rates, but the page provides no numeric interval bounds.
Context: This is an evaluator-reported whole-agent result, not an isolated model evaluation. The source page does not expose numeric confidence-interval bounds or an immutable API serving revision; no CI value is inferred from the evaluator's general interval method.

Source Cekura Bench speech-to-speech benchmark

Cekura all-check call pass rate (Show measurement, test conditions, and source)

Source value: 89
Score: 89
Metric: share of calls passing every configured check
Unit: %
Benchmark version: Cekura Bench GPT Realtime 2.1 speech-to-speech page, published and modified 2026-10-16; result-run revision not stated
Category: audio
Direction: Higher is better
Sample: 246
Source type: independent-evaluator
Evaluator: Cekura Bench speech-to-speech benchmark
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. Cekura labels the evaluated model GPT Realtime 2.1. This is the complete phone-agent configuration in a Pipecat pipeline, not a model-only score. The page reports that 89.0% of all 246 calls passed every configured check. The exact deployed agent commit and OpenAI API serving revision are not stated. The evaluator README describes scenario-stratified bootstrap intervals for rates, but the page provides no numeric interval bounds.
Context: This is an evaluator-reported whole-agent result, not an isolated model evaluation. The page does not publish numeric confidence-interval bounds or an immutable API serving revision.

Source Cekura Bench speech-to-speech benchmark

Cekura saved caller-data exactness (Show measurement, test conditions, and source)

Source value: 92
Score: 92
Metric: share of data-saving calls whose saved caller data was exactly correct
Unit: %
Benchmark version: Cekura Bench GPT Realtime 2.1 speech-to-speech page, published and modified 2026-10-16; result-run revision not stated
Category: audio
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Cekura Bench speech-to-speech benchmark
Status: active
Retrieved at: 2026-10-17
Methodology: Independent evaluation. Cekura labels the evaluated model GPT Realtime 2.1. This is a whole-agent data-handling outcome in the phone-agent benchmark. The page reports exact correctness on 92.0% of calls that save caller data, but does not publish the denominator for those calls. The exact deployed agent commit and OpenAI API serving revision are not stated.
Context: The denominator for calls that save data is not published, so sampleCount is omitted. This is a whole-agent result and not a model-only accuracy score.

Source Cekura Bench speech-to-speech benchmark

Cekura TurnBench end-of-turn score, Light noise (Show measurement, test conditions, and source)

Source value: 83
Score: 83
Metric: end-of-turn score for the named line condition, out of 100
Unit: %
Benchmark version: Cekura TurnBench leaderboard, last updated 2026-10-13
Direction: Higher is better
Sample: 36
Source type: independent-evaluator
Evaluator: Cekura TurnBench
Status: active
Retrieved at: 2026-10-23
Methodology: Independent evaluation. GPT Realtime 2.1; condition: Light noise. Cekura does not state a provider backend or configuration variant for this row.
Context: Cekura reports 36 calls per model and condition on a private leaderboard set of 25 scripts. This is a call count, not a count of 36 distinct independent tasks; the source does not publish an independent-task N or confidence interval.

Source Cekura TurnBench

Page 1 of 5

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 83 models

pyannote Community-173

Nemotron-3-Diarization57

Gemini 3.5 Transcribe52

Confucius4-R2T252

Granite Speech 5.0 470M TurboCTC52

Kokoro-82M51

Nova-351

Fish Audio S2 Pro51

NVIDIA Nemotron 3.5 Transcribe50

Granite Speech 5.0 470M TurboCTC NC50

Orukeet49

Lyria 3.547

Gemini 3.5 Live Translate46

Phonon-244

Aura-244

Gemini 3.1 Flash TTS44

MiniMax-Music341

Audio8-ASR-Infinite41

GPT-Realtime 2.141

Eleven v339

pyannote Community-1

Nemotron-3-Diarization

Gemini 3.5 Transcribe

Confucius4-R2T2

Granite Speech 5.0 470M TurboCTC

Kokoro-82M

Nova-3

Fish Audio S2 Pro

NVIDIA Nemotron 3.5 Transcribe

Granite Speech 5.0 470M TurboCTC NC

Orukeet

Lyria 3.5

Gemini 3.5 Live Translate

GPT-Realtime 2.1

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 5, 2026

GPT-Realtime 2.1Audio8-ASR-InfiniteMiniMax-Music3Eleven v3GPT-Live-Transcribe

All specifications

Architecture

Model class: Real-time audio model; Source: Model class
Knowledge cutoff: Sep 30, 2024; Source: Knowledge cutoff

Input and output

Input: Audio, image and text; Source: Input
Output: Audio and text; Source: Output
Context window: 128,000 tokens; Source: Context window
Maximum output: 32,000 tokens; Source: Maximum output
Reasoning: Configurable reasoning effort; tool calls supported; Source: Reasoning

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking: Subject to limitations. OpenAI lists an inaudible SynthID watermark for supported audio outputs through ChatGPT and the OpenAI API. Coverage can depend on the model, output path and creation date.; Source: AI watermarking
Watermarking technique: SynthID; Source: Watermarking technique
Labelling reviewed: Nov 5, 2026; Source: Labelling reviewed

Licence and access

Access: API; Source: Access

API

API model ID: gpt-realtime-2.1; Source: API model ID

Further details

Task: Voice dialogue; Source: Task
Supported details: Real-time dialogue with tool calls; Source: Supported details

Prices

Text tokens: $4 input, $0.40 cached input, $24 output per 1M tokensSource
Audio tokens: $32 input, $0.40 cached input, $64 output per 1M tokensSource
Image tokens: $5 input, $0.50 cached input per 1M tokensSource

Sources and data date

Show all 27 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

Type	Evidence and data date
Research status	Research date October 15, 2026. 21 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not published	per-model-evaluated-n Original wording The source reports 153k comparative ratings across 7 models, 819 prompts, and 12 questions overall, but does not state the number of ratings or completed cases for this model. No sampleCount is recorded.
Additional source	OpenAI model documentation (retrieved September 22, 2026; October 15, 2026) · OpenAI GPT-Realtime-2.1 model documentation · Editorial description reviewed October 2, 2026
Additional source	Artificial Analysis (retrieved October 15, 2026; October 28, 2026)1 2
Additional source	Cekura Bench speech-to-speech benchmark (retrieved October 17, 2026)
Additional source	David AI (retrieved October 18, 2026)
Additional source	VoxParity (retrieved October 20, 2026)
Additional source	Cekura TurnBench (retrieved October 23, 2026)
Additional source	DesignArena (retrieved October 28, 2026)1 2
Additional source	Speko (retrieved October 28, 2026)
Additional source	Model metadata (retrieved October 15, 2026; October 28, 2026)
Additional source	Model metadata (retrieved October 17, 2026)
Additional source	Model metadata (retrieved October 28, 2026)1 2 3 4 5 6 7 8
Additional source	OpenAI: Provenance signals (retrieved November 5, 2026)
Additional source	OpenRouter: Popularity index (derived), CC BY 4.0
Additional source	Vercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional source	Ollama: cumulative downloads for the popularity index
Additional source	Replicate: cumulative runs for the popularity index
Additional source	Google Trending Now: limited feed selection for Germany and the US
Additional source	Creative Commons Attribution 4.0