Soniox has launched Soniox STT v6, a proprietary speech-to-text API available as stt-rt-v6 for WebSocket streaming and stt-async-v6 for async REST across 60+ languages, priced at $0.12/hour streaming and $0.10/hour async with diarization and language ID included. The release introduces two key capabilities for voice agents: agent text input and model-level echo filtering. Agent text input, enabled via "enable_agent_text_input": true on stt-rt-v6 only, accepts the agent's spoken text as JSON frames {"type": "agent_text", "text": "...", "end_of_segment": true} streamed on the same WebSocket 500 ms to 1 second before playback, paced by TTS timestamps; on barge-in, played text is flushed with end_of_segment: true and unplayed text dropped. The model uses this context to recognize short caller replies — correcting errors like "River Roast" misheard as "River Road" — and to filter residual agent audio leaking past acoustic echo cancellation (AEC) without replacing it. The agent's text never appears in the transcript. Additional v6 improvements include better entity formatting for numbers, currencies, dates, and IDs; spoken commands like "period" rendered as words; enhanced meeting transcription, diarization, and code-switching handling. A formatting change writes "15 dollars and 20 cents" verbatim rather than "$15.20". No v6 benchmarks are published; v5 scored 1.34% semantic WER, 83.3% perfect transcripts, and 260 ms median TTFS on Pipecat. v5 models deprecate December 15, 2026, with automatic routing to v6. Agent text input is billed at $4.00 per 1M input tokens.