← Back to Voice models
Boson AI
Higgs Realtime
Released Aug 4, 2026
Boson AI's audio-native speech-to-speech model and hosted API for voice agents, compatible with the OpenAI Realtime API.
- Price per minute
- —
- Pricing
- Audio input $0.0023/min, audio output $0.014/min, transcription input $0.0025/min; underlying token rates $0.75/1M input, $0.25/1M cached, $4.50/1M output. Separate in/out per-minute rates, so no single perMinuteUsd.
- Inputs
- audio, text
- Tool calling
- Yes
- Languages
- 100
- Voices
- —
- Big Bench Audio
- 68.6%
- Access
- Generally available
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Audio at $0.0023/min in and $0.014/min out
- 100+ languages with code-switching mid-utterance
- Drop-in OpenAI Realtime API compatibility, including tool calling
Weaknesses
- Big Bench Audio 68.6% on AA, vs 94.8% in Boson's own run
- Its successor, Higgs Realtime Duplex (2026-11-05), is API-waitlist only
Not notable: #31 on AA Big Bench Audio (68.6%). Boson's blog reports BigBench Audio v1.2 at 94.8 (vs gpt-realtime-1.5 86.0), ComplexFuncBench (Audio) 83.4 and IHBench-RQ 71.2, evaluated 2026-07-02; the BBA figure conflicts with AA. Latency about 0.70s to first audio frame (Boson). Higgs Realtime Duplex, a full-duplex omni successor announced 2026-11-05, is API-waitlist only and not an entry.
Links
API model strings
- boson-ai
higgs-realtime