Audio modelActive
GPT-Live 1
OpenAI
- Released
- September 10, 2026
- Data date
- October 7, 2026
- Model class
- Access
GPT-Live 1 separates the voice conversation from reasoning. It can listen and speak at the same time while a separate backend model handles reasoning and tools.
OpenAI charges $0.05 per session minute, billed by the second. Backend model and tool use are billed separately, so the rate does not cover the full agent.
Specifications and access
| Specification | Value and source |
|---|---|
| Model class | Full-duplex voice modelSource |
| Access | OpenAI APISource |
| Input modalities | Text and audioSource |
| Output modalities | Text and audioSource |
| API model ID | gpt-live-1Source |
| Voice interaction | Can listen and speak at the same time; reasoning and tool use are delegated to a backendSource |
Page 1 of 4
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 (Sol low) (Show measurement, test conditions, and source)
- Source value
- 89
- Score
- 89
- Metric
- speech reasoning accuracy
- Unit
- %
- Benchmark version
- Big Bench Audio v1.2
- Category
- audio
- Direction
- Higher is better
- Sample
- 1000
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis Speech to Speech Index
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. GPT-Live-1 with GPT-Live backend Sol (low); 1,000 audio questions from four Big-Bench Hard categories using 23 synthetic voices; answers judged against the reference.
- Context
- The index report gives the percentage but does not break this model row down by question category.
Artificial Analysis Speech Agent Arena task success rate (Sol low) (Show measurement, test conditions, and source)
- Source value
- 90.9385
- Score
- 90.9385
- Metric
- eligible conversations with correct final task-completing tool call
- Unit
- %
- Benchmark version
- Current Speech Agent Arena result
- Category
- audio
- Direction
- Higher is better
- Sample
- 309
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis Speech Agent Arena
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. GPT-Live-1 with GPT-Live backend Sol (low); 281 successes and 28 model failures among eligible conversations. Participant deviations and unverifiable cases are excluded.
- Context
- Participant deviations and unverifiable conversations are excluded from the denominator; this is distinct from the τ-Voice result.
Artificial Analysis Big Bench Audio (Astra medium) (Show measurement, test conditions, and source)
- Source value
- 90.1
- Score
- 90.1
- Metric
- speech reasoning accuracy
- Unit
- %
- Benchmark version
- Big Bench Audio v1.2
- Category
- audio
- Direction
- Higher is better
- Sample
- 1000
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis Speech to Speech Index
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. GPT-Live-1 with GPT-Live backend GPT-6 Astra (medium); 1,000 audio questions from four Big-Bench Hard categories using 23 synthetic voices; answers judged against the reference.
- Context
- The index report gives the percentage but does not break this model row down by question category.
Artificial Analysis Speech Agent Arena task success rate (Astra medium) (Show measurement, test conditions, and source)
- Source value
- 87.3606
- Score
- 87.3606
- Metric
- eligible conversations with correct final task-completing tool call
- Unit
- %
- Benchmark version
- Current Speech Agent Arena result
- Category
- audio
- Direction
- Higher is better
- Sample
- 269
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis Speech Agent Arena
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. GPT-Live-1 with GPT-Live backend GPT-6 Astra (medium); 235 successes and 34 model failures among eligible conversations. Participant deviations and unverifiable cases are excluded.
- Context
- Participant deviations and unverifiable conversations are excluded from the denominator; this is distinct from the τ-Voice result.
OpenAI Tau3 end-to-end customer-service task success (Show measurement, test conditions, and source)
- Source value
- 83.6
- Score
- 83.6
- Metric
- Tau3 end-to-end customer-service task success
- Unit
- %
- Benchmark version
- OpenAI launch evaluation, 2026-09-10
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- OpenAI Developer Community GPT-Live-1 announcement
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. GPT-Live-1 paired with GPT-6 Astra at medium effort; first-attempt task success, equally weighted airline, retail, and telecom domains. OpenAI gives no sample count.
- Context
- Vendor-published result without independent replication; OpenAI discloses no sample count except for TauBanking.
OpenAI TauBanking task success (Show measurement, test conditions, and source)
- Source value
- 38.1
- Score
- 38.1
- Metric
- TauBanking task success
- Unit
- %
- Benchmark version
- OpenAI launch evaluation, 2026-09-10
- Category
- audio
- Direction
- Higher is better
- Sample
- 97
- Source type
- vendor-reported
- Evaluator
- OpenAI Developer Community GPT-Live-1 announcement
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. GPT-Live-1 paired with GPT-6 Astra at medium effort; 97 banking_knowledge tasks involving document retrieval and account tools.
- Context
- Vendor-published result without independent replication; OpenAI discloses no sample count except for TauBanking.
OpenAI Full Duplex Bench v1.5 interactivity (Show measurement, test conditions, and source)
- Source value
- 80.1
- Score
- 80.1
- Metric
- Full Duplex Bench v1.5 interactivity
- Unit
- %
- Benchmark version
- OpenAI launch evaluation, 2026-09-10
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- OpenAI Developer Community GPT-Live-1 announcement
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. GPT-Live-1 paired with GPT-6 Astra at medium effort; OpenAI's v1.5 interactivity benchmark. A per-run sample count is not supplied.
- Context
- Vendor-published result without independent replication; OpenAI discloses no sample count except for TauBanking.
OpenAI Full Duplex Bench v3 tool-calling success (Show measurement, test conditions, and source)
- Source value
- 87
- Score
- 87
- Metric
- Full Duplex Bench v3 tool-calling success
- Unit
- %
- Benchmark version
- OpenAI launch evaluation, 2026-09-10
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- OpenAI Developer Community GPT-Live-1 announcement
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. GPT-Live-1 with GPT Live backend Terra at low effort; separate Full Duplex Bench v3 spoken-request tool-use test.
- Context
- Vendor-published result without independent replication; OpenAI discloses no sample count except for TauBanking.
OpenAI Full Duplex Bench v3 response quality (Show measurement, test conditions, and source)
- Source value
- 90
- Score
- 90
- Metric
- Full Duplex Bench v3 response quality
- Unit
- %
- Benchmark version
- OpenAI launch evaluation, 2026-09-10
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- OpenAI Developer Community GPT-Live-1 announcement
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. GPT-Live-1 with GPT Live backend Terra at low effort; separate Full Duplex Bench v3 spoken answer quality test against reference intent.
- Context
- Vendor-published result without independent replication; OpenAI discloses no sample count except for TauBanking.
LiveLingo Interpreter Test English to Spanish translation coverage within 30 seconds (Show measurement, test conditions, and source)
- Source value
- 98
- Score
- 98
- Metric
- translation coverage within 30 seconds (percent)
- Unit
- %
- Benchmark version
- LiveLingo GPT-Live-1 Interpreter Test, measured 2026-09-15
- Category
- audio
- Direction
- Higher is better
- Sample
- 90
- Source type
- independent-evaluator
- Evaluator
- LiveLingo Research GPT-Live-1 interpreter test
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. gpt-live-1 API prompted as a simultaneous interpreter; 30 synthesized English sentences repeated three times with 0.5-second gaps into a 452.041-second real-time stream. LiveLingo aligns each emitted sentence to source timing; completion lag is last committed translation token minus source sentence end. Direction: English to Spanish; percentage of all 90 source clips receiving an aligned translation within 30 seconds.
- Context
- LiveLingo discloses that it builds a competing system included in the test. The published headline row is one 452-second run for this language pair; the 90 clips are 30 synthesized sentences repeated three times, not 90 independent utterances. The source audio is single-speaker synthesized speech without overlap or background noise. The test measures prompted interpreter timing and coverage, not translation fidelity or general voice-agent quality. Backend/reasoning settings are not specified.
LiveLingo Interpreter Test English to Japanese translation coverage within 30 seconds (Show measurement, test conditions, and source)
- Source value
- 17
- Score
- 17
- Metric
- translation coverage within 30 seconds (percent)
- Unit
- %
- Benchmark version
- LiveLingo GPT-Live-1 Interpreter Test, measured 2026-09-15
- Category
- audio
- Direction
- Higher is better
- Sample
- 90
- Source type
- independent-evaluator
- Evaluator
- LiveLingo Research GPT-Live-1 interpreter test
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. gpt-live-1 API prompted as a simultaneous interpreter; 30 synthesized English sentences repeated three times with 0.5-second gaps into a 452.041-second real-time stream. LiveLingo aligns each emitted sentence to source timing; completion lag is last committed translation token minus source sentence end. Direction: English to Japanese; percentage of all 90 source clips receiving an aligned translation within 30 seconds.
- Context
- LiveLingo discloses that it builds a competing system included in the test. The published headline row is one 452-second run for this language pair; the 90 clips are 30 synthesized sentences repeated three times, not 90 independent utterances. The source audio is single-speaker synthesized speech without overlap or background noise. The test measures prompted interpreter timing and coverage, not translation fidelity or general voice-agent quality. Backend/reasoning settings are not specified. For English to Japanese, the report deliberately withholds headline lag because 20 of 90 source sentences never arrive; only coverage is recorded here.
Artificial Analysis Speech to Speech Index v2.0 (Sol low) (Show measurement, test conditions, and source)
- Source value
- 80.0805683287912
- Score
- 80.0805683287912
- Metric
- equal-weight composite score on a 0–100 scale
- Unit
- points
- Benchmark version
- Index v2.0; current from August 2026
- Category
- audio
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Artificial Analysis Speech to Speech Index
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Independent evaluation. GPT-Live-1 with GPT-Live backend Sol (low); equal-weight composite of Big Bench Audio, τ-Voice, Arena Score, and Task Success Rate. Conversational Dynamics is not part of the current index.
- Context
- This result belongs to the complete voice-agent stack including the selected reasoning backend; it is not a frontend-only model score.
Page 1 of 4
Published prices
- Session price
- $0.05 per voice-session minute, billed per second. Backend model and tools cost extra.Source
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 7, 2026. 8 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | backend-independent-result Original wording Published performance results apply to Sol low, Astra medium, or Terra low. No benchmark isolates GPT-Live-1 without a named backend. |
| Unresolved | tau-voice-trial-count Original wording AA's general method describes three τ-Voice trials, but its current dashboard explicitly marks both GPT-Live-1 backend rows as one-trial results. |
| Not published | openai-sample-counts Original wording OpenAI gives a sample count of 97 only for TauBanking. Sample sizes are absent for Tau3, Full Duplex Bench v1.5, the v3 metrics, and reported turn-taking latency. |
| Unresolved | arena-appearance-reconciliation-sol Original wording The scenario slices sum to 634 appearances against 652 overall. The source does not explain the difference of 18 appearances. |
| Unresolved | arena-appearance-reconciliation-astra Original wording The scenario slices sum to 598 appearances against 610 overall. The source does not explain the difference of 12 appearances. |
| Additional source | OpenAI GPT-Live 1 model documentation (retrieved October 7, 2026) · OpenAI GPT-Live 1 model documentation · Editorial description reviewed October 7, 2026 |
| Additional source | OpenAI GPT-Live 1 launch and pricing (retrieved October 7, 2026) · OpenAI GPT-Live 1 launch and pricing · Editorial description reviewed October 7, 2026 |
| Additional source | Artificial Analysis Speech to Speech Index (retrieved October 7, 2026) |
| Additional source | Artificial Analysis Speech Agent Arena (retrieved October 7, 2026) |
| Additional source | OpenAI Developer Community GPT-Live-1 announcement (retrieved October 7, 2026) |
| Additional source | LiveLingo Research GPT-Live-1 interpreter test (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |
| Additional source | Model metadata (retrieved October 7, 2026) |