Amazon Web ServicesAudio modelActive
Amazon Nova 2 Sonic
n/ano search or usage data
Summary
Amazon Nova 2 Sonic is Bedrock’s speech-to-speech model for bidirectional real-time conversations. It accepts audio and text and returns both. Turn taking and expressive polyglot voices belong to the conversation model, while asynchronous tools run alongside the audio exchange.
AWS cites a one-million-token context and availability in US regions and Tokyo. Audio and text tokens are billed separately at the applicable Bedrock rate. It is a hosted Bedrock service rather than a local weights checkpoint.
On this page
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 71 published benchmark results
Page 1 of 6
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.
Big Bench Audio accuracy (Show measurement, test conditions, and source)
- Source value
- 87
- Metric
- accuracy
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 12; Big Bench Audio version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. Artificial Analysis evaluation of speech-to-speech models
- Context
- Artificial Analysis is named in Amazon's report; this record preserves Amazon's report rather than an independently retrieved leaderboard run.
BFCL accuracy (Show measurement, test conditions, and source)
- Source value
- 74.5
- Metric
- accuracy
- Unit
- %
- Benchmark version
- BFCL-v3 filtered snapshot, repository commit c67d246
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. single-turn, single-function, python-only subset; text prompts transformed with Google TTS; report explicitly cites BFCL state as of 2025-03-02
- Context
- The cited 2025-03-02 BFCL snapshot is a dataset state, not a Nova 2 Sonic run date.
ComplexFunction accuracy (Show measurement, test conditions, and source)
- Source value
- 65.2
- Metric
- accuracy
- Unit
- %
- Benchmark version
- ComplexFunction; version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. text prompts transformed with Google TTS
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
IFBench prompt-level accuracy (Show measurement, test conditions, and source)
- Source value
- 33.3
- Metric
- accuracy
- Unit
- %
- Benchmark version
- IFBench; version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. text prompts transformed with Google TTS
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
IFBench instruction-level accuracy (Show measurement, test conditions, and source)
- Source value
- 37.5
- Metric
- accuracy
- Unit
- %
- Benchmark version
- IFBench; version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. text prompts transformed with Google TTS
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, American English masculine, versus GPT-Realtime (Show measurement, test conditions, and source)
- Source value
- 53.9
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, American English masculine, versus Gemini 2.5 Flash Live (Show measurement, test conditions, and source)
- Source value
- 60
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, Indian English masculine, versus GPT-Realtime (Show measurement, test conditions, and source)
- Source value
- 55.5
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, Indian English masculine, versus Gemini 2.5 Flash Live (Show measurement, test conditions, and source)
- Source value
- 64.2
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, Spanish masculine, versus GPT-Realtime (Show measurement, test conditions, and source)
- Source value
- 68.4
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, Spanish masculine, versus Gemini 2.5 Flash Live (Show measurement, test conditions, and source)
- Source value
- 70.3
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Human preference, French masculine, versus GPT-Realtime (Show measurement, test conditions, and source)
- Source value
- 54.7
- Metric
- win rate
- Unit
- %
- Benchmark version
- Nova 2 technical report Table 11; preference test version not stated
- Category
- audio
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Amazon Nova 2 technical report
- Status
- active
- Retrieved at
- 2026-10-04
- Methodology
- Vendor-reported result. human evaluation; rows are winning rates for Nova 2 Sonic against GPT-Realtime / Gemini 2.5 Flash Live
- Context
- Source-specific observation; it does not establish a compatible cross-model cohort.
Page 1 of 6
Similar models
More models of the same kind as Amazon Nova 2 Sonic, chosen by model line, developer, size, price and release date.
- MiniMax Speech 2.8MiniMax · released around the same time
- GPT-Live-1.1OpenAI · released around the same time
- MAI-Transcribe-1.5Microsoft · released around the same time
- MAI-Voice-2Microsoft · released around the same time
- Fish Audio S2.1 ProFish Audio · released around the same time
- GPT-Realtime 2.1 MiniOpenAI · released around the same time
- Cartesia Ink-2Cartesia · released around the same time
- Fish Audio S2 ProFish Audio · released around the same time
All specifications
Architecture
Input and output
- Capabilities
- Expressive polyglot voices, turn-taking control, and asynchronous tools
- 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
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- Shutdown
- None announced
- no source
API
Further details
Prices
- Billing
- Separate token rates for audio and text input and audio and text output; text rates also apply to transcription, tool calls, grounding, and history. Standard, Priority, and Flex tiers must be checked separately.Source
Sources and data date
Show all 10 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.