GoogleAudio modelActive
Gemini 3.8 Live
PopularityRank 663 of 875
0out of 100same as last month
Summary
Gemini 3.8 Live is designed for direct audio responses without an extra reasoning delay. It accepts text, images, audio, and video and returns text and audio. While an asynchronous function call runs, the session can keep streaming audio and updating session content; search grounding and built-in audio streaming are part of its documented live behavior.
The API profile allows 131,072 input tokens and 65,536 output tokens. Gemini 3.8 Live Extended Thinking is the separate variant for background reasoning, while Gemini 3.8 Live keeps direct dialogue and ongoing tool use open without that extra step.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 143 published benchmark results
Page 1 of 12
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
- 92
- Score
- 92
- 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: Gemini 3.8 Live.
Artificial Analysis Conversational Dynamics (Show measurement, test conditions, and source)
- Source value
- 96.1
- Score
- 96.1
- 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: Gemini 3.8 Live.
Artificial Analysis τ-Voice task completion (Show measurement, test conditions, and source)
- Source value
- 30.1
- Score
- 30.1
- 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 Gemini 3.8 Live uses 3 trials.
Speech Agent Arena task success rate (Show measurement, test conditions, and source)
- Source value
- 93.2
- Score
- 93.2
- 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
- 89
- Upper confidence bound
- 95.8
- 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: Gemini 3.8 Live.
Speko S2S concierge call completion (Show measurement, test conditions, and source)
- Source value
- 98
- Score
- 98
- Metric
- scripted call completion rate
- Unit
- %
- Benchmark version
- Speko S2S board v2; exact route google:gemini-3.8-live; measured 2026-09-16
- Category
- speech-to-speech / concierge task quality
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Speko speech-to-speech concierge benchmark
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact Speko route google:gemini-3.8-live, displayed model gemini-3.8-live; six multi-turn concierge scenarios, three sessions per scenario (18 sessions per model); scripted call completion rate. Measured 2026-09-16; evaluator states models were tested in their production-stack configuration.
- Context
- The evaluator publishes six scenarios and three sessions each, but not the denominator or confidence interval for each metric. No immutable backend revision or all request settings are listed. Latency and pricing cells are excluded.
Cekura Gemini 3.8 Live scenario reliability (pass³) (Show measurement, test conditions, and source)
- Source value
- 41.5
- Score
- 41.5
- Metric
- share of scenarios passing on all three runs (pass^3)
- Unit
- %
- Benchmark version
- Cekura Gemini Live benchmark, page dated 2026-10-16
- Category
- speech-to-speech
- Direction
- Higher is better
- Sample
- 82
- Source type
- independent-evaluator
- Evaluator
- Cekura speech-to-speech benchmark
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Cekura's exact row is Gemini 3.8 Live, with no thinking level. The benchmark uses the same 82 scenarios for every model, each run three times, for 246 live calls per model. Reliability is the pass^3 scenario rate; the page reports 41.5%.
- Context
- The API serving revision is not pinned and the page does not publish a confidence interval. Do not confuse scenario N=82 with 246 calls.
Cekura TurnBench end-of-turn score (six-condition average) (Show measurement, test conditions, and source)
- Source value
- 82
- Score
- 82
- Metric
- end-of-turn score averaged across six line conditions
- Unit
- %
- Benchmark version
- Cekura TurnBench leaderboard, retrieved 2026-10-17
- Category
- speech-to-speech / turn-taking
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- Cekura TurnBench
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. The exact leaderboard row is Gemini 3.8 Live. The source reports the end-of-turn score from Sesame TurnBench's published scorer, averaged over six line conditions and a 25-script private leaderboard set. It reports 36 calls per model per condition; all models heard the same recorded calls, streamed from US East with provider-default turn detection.
- Context
- The private benchmark set cannot be rerun externally. The page says its 95% range overlaps with several models but gives no exact interval for this row; no CI or serving revision is recorded.
VAmoS Energy verified task completion (Show measurement, test conditions, and source)
- Source value
- 40.1
- Score
- 40.1
- Metric
- completion rate, fraction of counted attempts passing all task assertions
- Unit
- %
- Benchmark version
- VAmoS Part Deux v1, arXiv:2609.38512 (2026-09-29)
- Category
- speech-to-speech / voice-agent task completion
- Direction
- Higher is better
- Sample
- 284
- Lower confidence bound
- 34.6
- Upper confidence bound
- 45.9
- Source type
- independent-evaluator
- Evaluator
- VAmoS Part Deux, VAmoS Energy
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. The paper's Table 2 reports the exact speech-to-speech stack Gemini 3.8 Live, configured as gemini-3.8-live with the same bridge as Gemini 3.1 Live. VAmoS Energy uses 100 household utility-billing tasks, 16 tools backed by a stateful Stripe billing twin and Apache Fineract, two to four requests per call, and three repeats per task across 14 stacks. The completion denominator counts 284 retained Gemini attempts after benchmark-fault exclusions; agent-side run failures count as misses.
- Context
- This is the whole voice-agent stack, not an isolated model-only test. The benchmark globally excluded 402 fault-affected calls; the per-model retained attempt count is the reported n=284.
MPTB STOP yield on successful interruptions (Show measurement, test conditions, and source)
- Source value
- 100
- Score
- 100
- Metric
- Yield rate on genuine floor-acquisition interruptions
- Unit
- %
- Benchmark version
- MultiPartyTurnBench, arXiv:2610.20172v1, Table 4
- Category
- speech-interaction
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- MultiPartyTurnBench
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact paper label Gemini-3.8-Live (Google, 2026); full-duplex agent test on 120 interactive floor-control scenarios repeated three times per system. Appendix G.3 attributes every stop on true interruptions to provider barge-in events, not solely to model reasoning.
- Context
- Whole-agent behavior, with provider VAD cancellation; not a model-only reasoning score. The plain live model was tested, not extended thinking. Serving revision is not published.
MPTB STOP false yield on backchannels (Show measurement, test conditions, and source)
- Source value
- 97.7
- Score
- 97.7
- Metric
- False-yield rate on backchannels that do not acquire the floor
- Unit
- %
- Benchmark version
- MultiPartyTurnBench, arXiv:2610.20172v1, Table 4 and Appendix G.1
- Category
- speech-interaction
- Direction
- Lower is better
- Sample
- 88
- Source type
- independent-evaluator
- Evaluator
- MultiPartyTurnBench
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact paper label Gemini-3.8-Live (Google, 2026); STOP test on non-floor-taking backchannels. Appendix G.1 gives 86 yields out of 88 backchannels; all tested Gemini stops coincide with provider barge-in cancellation.
- Context
- The observed stops are tied to provider VAD/barge-in behavior, not a pure model score. The denominator is 88 backchannel events, not independent scenarios. Serving revision is not published.
MPTB STOP balanced accuracy (Show measurement, test conditions, and source)
- Source value
- 51.1
- Score
- 51.1
- Metric
- Balanced accuracy for selective interruption yielding; 50% is chance
- Unit
- %
- Benchmark version
- MultiPartyTurnBench, arXiv:2610.20172v1, Table 4
- Category
- speech-interaction
- Direction
- Higher is better
- Lower confidence bound
- 50
- Upper confidence bound
- 52.8
- Source type
- independent-evaluator
- Evaluator
- MultiPartyTurnBench
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Independent evaluation. Exact paper label Gemini-3.8-Live (Google, 2026); STOP balanced accuracy over true interruptions and backchannels. Table 4 brackets are 95% scenario-clustered bootstrap confidence intervals. Appendix G.3 finds stops coincide with provider VAD barge-in events.
- Context
- Whole-agent behavior with provider VAD cancellation, not a model-only score. The CI is scenario-clustered. The paper tests plain Gemini-3.8-Live, not extended thinking. Serving revision is not published.
IndicFDB English (Indian) pause handling success rate (Show measurement, test conditions, and source)
- Source value
- 100
- Score
- 100
- Metric
- pause handling success rate
- Unit
- %
- Benchmark version
- IndicFDB arXiv:2609.31967v1, Table 2
- Category
- speech-to-speech / full-duplex
- Direction
- Higher is better
- Sample
- 315
- Source type
- community-reported
- Evaluator
- IndicFDB paper, arXiv:2609.31967v1
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. IndicFDB per-language full-duplex evaluation. Exact source row: Gemini 3.8 Live, cited official model ID gemini-3.8-live. Language: English (Indian). The paper does not report a pinned serving revision.
- Context
- Primary paper table; institutional independence is not established. Authors list NVIDIA and VoiceArena affiliations. No 95% confidence interval or pinned serving revision is published. Results are language-specific and are not pooled across languages. N=315 pause-handling samples per language; S.R. is the category mean over all samples. Method limitation: IndicFDB uses alignment-free, multi-threshold VAD heuristics. These may mislabel corner cases (for example, brief answers as backchannels); transcription and translation errors can accumulate before judging.
Page 1 of 12
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-1, pyannote: 73
Nemotron-3-Diarization, NVIDIA: 57
Gemini 3.5 Transcribe, Google: 52
Confucius4-R2T2, NetEase Youdao: 52
Granite Speech 5.0 470M TurboCTC, IBM: 52
Kokoro-82M, hexgrad: 51
Nova-3, Deepgram: 51
Fish Audio S2 Pro, Fish Audio: 51
NVIDIA Nemotron 3.5 Transcribe, NVIDIA: 50
Granite Speech 5.0 470M TurboCTC NC, IBM: 50
Orukeet, Oruk: 49
Lyria 3.5, Google: 47
Gemini 3.5 Live Translate, Google: 46
Phonon-2, Fermion Research: 44
Aura-2, Deepgram: 44
Gemini 3.1 Flash TTS, Google: 44
MiniMax-Music3, MiniMax: 41
Audio8-ASR-Infinite, Edge0 AI: 41
GPT-Realtime 2.1, OpenAI: 41
Gemini 3.8 Live, Google: 0
Gemini 3.8 Live
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 5, 2026
Gemini 3.8 LiveGPT-Live 1Suno v6YuE2-3BAmazon Nova 2 Sonic
All specifications
Architecture
Input and output
- Capabilities
- Interleaved reasoning, asynchronous function calling, session content updates, and built-in audio streaming
- Source: Capabilities
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
- Yes. Google documents SynthID for audio outputs of the listed Gemini Audio models.
- Source: AI watermarking
Licence and access
API
Further details
Prices
- Text input price
- $0.75 per 1M text tokensSource
- Audio input price
- $3 per 1M audio tokens ($0.005 per audio minute)Source
- Text output price
- $4.50 per 1M text tokensSource
- Audio output price
- $12 per 1M audio tokens ($0.018 per audio minute)Source
- Pricing basis
- Paid Standard tier, prices in USD. Per-minute rates are estimates based on token usage, not fixed per-minute charges.Source
Sources and data date
Show all 31 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. 24 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-metric-denominator-and-serving-revision Note No per-metric denominator, confidence interval or immutable backend revision is published. |
| Not published | Gemini Live TTFA sample count and serving revision Note The source does not provide per-model N, confidence interval, or a pinned serving revision for TTFA. |
| Not published | isolated-model-result Note VAmoS measures the full agent stack. It does not provide an isolated model-only result. |
| Unresolved | Coval rolling 30-day speech-to-speech measurements Note The exact public Coval page for Gemini 3.8 Live reports 0 samples / 30 days and says no dated, qualifying measurements are available in this window. HTTP Date Wed, 28 Oct 2026 00:07:14 GMT, ETag "16ypny4lbmn1uum"; no score or run time is published. The response timestamp is older than on the other inspected Coval pages, so current availability remains unresolved. |
| Additional source | Google Gemini model documentation (retrieved September 24, 2026; October 15, 2026; October 17, 2026) · Google Gemini 3.8 Live model documentation · Editorial description reviewed October 2, 2026 |
| Additional source | Official API pricing (retrieved October 25, 2026) |
| Additional source | Artificial Analysis (retrieved October 15, 2026; October 17, 2026; October 28, 2026) |
| Additional source | Artificial Analysis (retrieved October 15, 2026; October 28, 2026) |
| Additional source | Speko speech-to-speech concierge benchmark (retrieved October 17, 2026; October 28, 2026) |
| Additional source | Cekura speech-to-speech benchmark (retrieved October 17, 2026; October 23, 2026) |
| Additional source | Cekura TurnBench (retrieved October 17, 2026; October 23, 2026) |
| Additional source | VAmoS Part Deux, VAmoS Energy (retrieved October 17, 2026) |
| Additional source | MultiPartyTurnBench (retrieved October 17, 2026) |
| Additional source | VoxParity benchmark README, full-bank leaderboard (retrieved October 17, 2026) |
| Additional source | IndicFDB paper, arXiv:2609.31967v1 (retrieved October 18, 2026) |
| Additional source | Omi Health Medical Speech-to-Text Benchmark (retrieved October 18, 2026) |
| Additional source | Bluejay Labs, MIVAS leaderboard (retrieved October 22, 2026) |
| Additional source | Model metadata (retrieved October 15, 2026; October 28, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)12 |
| Additional source | Model metadata (retrieved October 28, 2026)12345678 |
| Additional source | Google DeepMind Gemini Audio (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 |