Models

Aethex runs three models in sequence on every call:

• Kora Read — Aethex's speech-to-text model. Converts the caller's audio to text in real time.
• LLM — the language model that generates the agent's responses. Configurable per agent.
• Kora Speak — Aethex's text-to-speech model. Converts the LLM's text response back to audio.

STT and TTS are always Aethex's own models. The LLM is the only configurable stage.

Kora Read is Aethex's ASR model. It is optimized for telephone-quality audio, handles regional accents, and maintains low latency with streaming transcription.

Kora Read is always active; it cannot be swapped for a third-party STT. If you need higher accuracy on a specific accent or language, contact our team — fine-tuned checkpoints are available on request.

Kora Speak is Aethex's neural TTS model. It produces natural-sounding speech with controllable prosody, supports multiple languages, and streams audio in real time so the first syllable arrives in under 200ms.

Select the voice by setting `voice_id` on the agent. Browse available voices on the Voices concept page.

The llm_model field on the agent selects the language model. Set it to "aethex-default" (or an empty string "") to use the Aethex default (recommended for most builds). For other providers, use a flat public model name from GET /api/v1/models, such as claude-sonnet-4-6 or gpt-4o.

Support for additional providers and models is added continuously. Contact our team if you need a model not listed here.

Use PATCH /agents/{id} with llm_model to switch models. Here are the four most common options:

# Aethex default (recommended for first builds)
requests.patch(f"{BASE_URL}/agents/{AGENT_ID}", headers=headers, json={
    "llm_model": "aethex-default",
})

# OpenAI
requests.patch(f"{BASE_URL}/agents/{AGENT_ID}", headers=headers, json={
    "llm_model": "gpt-4o",
})

# Anthropic
requests.patch(f"{BASE_URL}/agents/{AGENT_ID}", headers=headers, json={
    "llm_model": "claude-sonnet-4-6",
})

# Google
requests.patch(f"{BASE_URL}/agents/{AGENT_ID}", headers=headers, json={
    "llm_model": "gemini-2.5-pro",
})

# Use the flat public names from GET /api/v1/models. Legacy "openai:" /
# "anthropic:" / "provider/slug" values still work but are deprecated, and
# agent responses always return the public name — never the internal route.

# Both "aethex-default" and an empty string "" select the platform default.
# Send "" to explicitly move an agent back to the default: because a response
# already shows "aethex-default" for a defaulted agent, PATCHing that same
# value back is a no-op, so "" is the unambiguous reset.
requests.patch(f"{BASE_URL}/agents/{AGENT_ID}", headers=headers, json={
    "llm_model": "",
})

Every additional network hop to a third-party LLM adds latency. The Aethex default model is co-located with STT and TTS, so the round trip is sub-100ms. Cloud LLMs typically add 200–800ms depending on provider and model size.

For latency-sensitive applications (fast-paced conversations, real-time customer service), start with the Aethex default. Move to a third-party model only when you need capabilities that the default does not provide.