Models

An agent names its models as {provider, model} pairs. The only valid pairs are the ones in the provider catalog, and a config naming anything else is refused on save with the list of what it could have said instead. The catalog is live. GET /v1/catalog is authoritative; the tables at the bottom of this page are a snapshot of it and can lag behind.

curl https://api.in.talqing.com/v1/catalog \
  -H "Authorization: Bearer $TALQING_API_KEY"

const catalog = await talqing.catalog.get();

for (const entry of catalog.llm) {
  console.log(entry.provider, entry.model, entry.reasoning_efforts, entry.vision);
}

catalog = talqing.catalog.get()

for entry in catalog["llm"]:
    print(entry["provider"], entry["model"], entry["reasoning_efforts"], entry["vision"])

The response carries one array per kind — llm, stt, tts, realtime, avatar, noise_cancellation — plus providers, the union of every language an agent may be set to, the subset of those the turn detector understands, system_vars, and the platform fee per channel.

Every model runs on your provider key. An agent will not publish until a key is configured for each provider it needs — the language model, the speech models, the avatar, the noise canceller, and any fallback among them. See provider keys.

Two pipelines

A voice or video agent runs one of two shapes, and config.realtime is what discriminates them. Set it, and it is the whole pipeline; leave it null, and the cascade runs.

	Cascade (`stt` → `llm` → `tts`)	Realtime (speech-to-speech)
Models	three, mixed freely across providers	one
Latency	three network hops per turn, one of which is the model	one socket, lowest time to first sound
Turn detection	ours, from the speech-to-text model and the language	the provider’s, inside its own socket
Voice	any voice from any text-to-speech provider	the model’s own roster
Fixed wording	the greeting, a `say` operation and a handoff message are spoken word for word	delivered as “say this” instructions the model may reword
`reasoning_effort`, `priority`, `builtin_tools`	yes	no
`fallback`	yes, per stage	no
Preemptive generation	yes	no, pinned off
`on_user_turn_completed` hook	yes	refused
Screen share	yes	refused
Expressive delivery	on the two models that speak tags	no
Recording consent notice	yes	refused
Images people attach	if the model reads images	if the model reads images — the two `grok-voice-*` entries do not

Take the cascade unless latency is the product. It is the shape everything else on the platform is built around: exact wording, per-stage failover, provider-side tools, a screen the agent can watch, and one hook that fires on every user turn. What you pay for it is a longer path to the first syllable. Take a realtime model when the conversation is the interface — a demo, a companion, an interview — and when you can live with a greeting that is approximate and a call that has no failover. A realtime agent still records, still analyses, still runs tools, knowledge, MCP servers and handoffs. Switching between them is a normalization, not an error: setting realtime clears stt, llm and tts, and setting them back means also setting realtime to null. See channels.

The fields on each slot

Slot	Fields
`stt`	`provider`, `model`, `fallback`
`llm`	`provider`, `model`, `reasoning_effort`, `priority`, `builtin_tools`, `fallback`
`tts`	`provider`, `model`, `voice`, `speed`, `expressive`, `stability`, `similarity_boost`, `fallback`
`realtime`	`provider`, `model`, `voice`, `speed`

voice_name also exists on tts and realtime. It is a display label for the dashboard and is ignored when the agent is compiled. speed accepts 0.5 to 2.0 as a number, and is then checked against the chosen model’s own range — the tables below carry it. A model with no speed control refuses anything but 1.0. Leaving voice null uses the entry’s default voice. stability and similarity_boost belong to the ElevenLabs voice you picked and are refused on every other provider; see voices.

Choosing

The language is the first constraint, not the last. Pick the language, then models that cover it — a model that publishes languages and does not speak the one you chose is a save error, and a model with language_required refuses Auto outright. See language. After that:

Speech-to-text is the biggest lever on how the agent feels. A streaming model ends the turn on its own detector; a batch model transcribes each utterance in one request, which is cheaper and adds roughly half a second before the agent replies. Word-count interruption also stops working on a batch model, because there are no interim transcripts. See speech-to-text and turn detection.
The language model is where thinking time is spent. On a call, every second of reasoning is silence — take the least the model will accept. See thinking and provider tools.
Text-to-speech is where the agent’s character is. Price per character ranges over a factor of twenty-five across this catalog, and the two models that speak delivery tags are the only ones expressive delivery works on.
Mixing providers is normal. The three cascade slots are independent, and the agent’s single language is translated into each provider’s own spelling for you.

The priority lane

llm.priority asks the provider to schedule this model’s turns ahead of standard traffic, which steadies the worst case rather than speeding up the average. It costs between 1.75x and 2.5x the standard token price depending on the model, so read the priority lane before switching it on.

Reference

A snapshot of the catalog. GET /v1/catalog wins if the two disagree. Prices are USD and are what the provider bills you on your own key — Talqing adds only the per-minute platform fee.

Language models

Every entry below runs on text, voice and video, and every one of them reads an image. “Thinking” is the model’s reasoning_efforts, ordered fastest first; the first value is what an agent gets until its author picks another, and an empty cell means the model has no such setting.

Provider	Model	Thinking efforts	Provider tools	Priority lane	Price per 1M tokens (in / cached in / out)
xai	`grok-4.5`	minimal, low, medium, high	`web_search`, `x_search`, `code_interpreter`, `file_search`	yes	$2.00 /$ 0.30 / $6.00
xai	`grok-4.3`	none, minimal, low, medium, high	`web_search`, `x_search`, `code_interpreter`, `file_search`	yes	$1.25 /$ 0.20 / $2.50
openai	`gpt-5.6-sol`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$5.00 /$ 0.50 / $30.00
openai	`gpt-5.6-terra`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$2.00 /$ 0.20 / $12.00
openai	`gpt-5.6-luna`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$0.20 /$ 0.02 / $1.20
openai	`gpt-5.5`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$5.00 /$ 0.50 / $30.00
openai	`gpt-5.4`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$2.50 /$ 0.25 / $15.00
openai	`gpt-5.4-mini`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	yes	$0.75 /$ 0.075 / $4.50
openai	`gpt-5.4-nano`	none, low, medium, high	`web_search`, `code_interpreter`, `file_search`	no	$0.20 /$ 0.02 / $1.25
openai	`gpt-4.1`	—	`web_search`, `code_interpreter`, `file_search`	yes	$2.00 /$ 0.50 / $8.00
openai	`gpt-4.1-mini`	—	`web_search`, `code_interpreter`, `file_search`	yes	$0.40 /$ 0.10 / $1.60
gemini	`gemini-3.6-flash`	minimal, low, medium, high	`google_search`, `url_context`, `code_execution`	yes	$1.50 /$ 0.15 / $7.50
gemini	`gemini-3.5-flash`	minimal, low, medium, high	`google_search`, `url_context`, `code_execution`	yes	$1.50 /$ 0.15 / $9.00
gemini	`gemini-3.5-flash-lite`	minimal, low, medium, high	`google_search`, `url_context`, `code_execution`	yes	$0.30 /$ 0.03 / $2.50
gemini	`gemini-3.1-flash-lite`	minimal, low, medium, high	`google_search`, `url_context`, `code_execution`	yes	$0.25 /$ 0.025 / $1.50

The default for a new agent is openai/gpt-5.6-luna.

Speech-to-text

Voice and video only. “On Auto” is what happens when config.language is null. Language counts are the codes the entry publishes; the codes themselves come back from the catalog endpoint.

Provider	Model	Streaming	On Auto	Languages	Price per audio second
xai	`xai-stt`	yes	detects (`en` is sent, and only affects number formatting)	25	$0.0000556
xai	`xai-stt-batch`	no	detects (`en` is sent, and only affects number formatting)	25	$0.0000278
sarvam	`saaras:v3`	yes	detects	23	$0.0000833
raya	`bakbak-stt`	no	no detection — transcribes as English	23	$0.000065
soniox	`stt-rt-v5`	yes	detects	60	$0.0000333
elevenlabs	`scribe_v2_realtime`	yes	detects	19	$0.0001084
elevenlabs	`scribe_v2`	no	detects	19	$0.0000611
openai	`gpt-realtime-whisper`	yes	detects	57	$0.0002833
openai	`gpt-4o-transcribe`	no	detects	57	$0.0001
deepgram	`nova-3-general`	yes	detects and code-switches	35	$0.0000967
deepgram	`nova-3`	yes	no detection — transcribes as English	35	$0.00008
deepgram	`flux-general-en`	yes	English only	1	$0.0001083
deepgram	`flux-general-multi`	yes	detects	35	$0.00013

Two entries detect nothing and transcribe every caller as English on Auto: raya/bakbak-stt and deepgram/nova-3. On Raya the failure is not symmetric — English audio read as Hindi comes back transliterated with its meaning intact, but Hindi audio read as English comes back with words missing, as a successful transcription with nothing to detect it by. If the agent takes non-English calls on either of these, set config.language rather than relying on Auto.

A streaming model bills the open socket for the whole call; a batch model bills only the utterances it was handed, which is roughly the caller’s share of the conversation. The default for a new agent is sarvam/saaras:v3.

Text-to-speech

Voice and video only. Speed is the model’s own accepted range.

Provider	Model	Languages	Auto	Speed	Voice settings	Expressive tags	Price
xai	`xai-tts`	21	yes	0.7 – 1.5	no	yes	$0.000015 / char
sarvam	`bulbul:v3`	11	yes (the model takes the language from the text)	0.5 – 2.0	no	no	$0.00003 / char
raya	`m1`	13	yes, speaks `en`	0.5 – 1.5	no	no	$0.0000039 / char
raya	`standard`	13	yes, speaks `en`	0.5 – 1.5	no	no	$0.0000039 / char
soniox	`tts-rt-v2`	63	no — a language is required	0.7 – 1.3	no	no	$0.0000141 / char
elevenlabs	`eleven_turbo_v2_5`	32	yes	0.8 – 1.2	yes	no	$0.00005 / char
elevenlabs	`eleven_flash_v2_5`	32	yes	0.8 – 1.2	yes	no	$0.00005 / char
elevenlabs	`eleven_v3`	33	yes	no speed control	yes	yes	$0.00010 / char
openai	`gpt-4o-mini-tts`	not language-aware	n/a	0.5 – 2.0	no	no	$0.60 in /$ 12.00 out per 1M tokens
openai	`tts-1`	not language-aware	n/a	0.5 – 2.0	no	no	$0.000015 / char
openai	`tts-1-hd`	not language-aware	n/a	0.5 – 2.0	no	no	$0.00003 / char
deepgram	`aura-2`	encoded in the voice	n/a	no speed control	no	no	$0.00003 / char
deepgram	`aura-1`	encoded in the voice	n/a	no speed control	no	no	$0.000015 / char

“Voice settings” means the model reads the stability and similarity_boost of the voice you picked. Only ElevenLabs sells them, and sending them anywhere else is a save error. “Expressive tags” means the model has a delivery-tag dialect the agent can be taught — xai/xai-tts and elevenlabs/eleven_v3 are the only two, and their dialects are different. The default for a new agent is sarvam/bulbul:v3 with the voice shubh. Voices are browsed separately, per provider and model, because most rosters are fetched live from the provider. See voices.

Realtime (speech-to-speech)

Voice and video only. One of these replaces stt, llm and tts together.

Provider	Model	Reads images	Languages	Speed	Voice roster
openai	`gpt-realtime-2.1`	yes	57, detects on Auto	0.5 – 2.0	10 voices, default `marin`
openai	`gpt-realtime-2.1-mini`	yes	57, detects on Auto	0.5 – 2.0	10 voices, default `marin`
gemini	`gemini-2.5-flash-native-audio-preview-12-2025`	yes	19, code-switches on its own	no speed control	30 voices, default `Puck`
xai	`grok-voice-think-fast-2.0`	no	steered from the prompt only	no speed control	fetched live, default `ara`
xai	`grok-voice-think-fast-1.0`	no	steered from the prompt only	no speed control	fetched live, default `ara`

Pricing, USD per 1M tokens except where noted:

Model	Text in	Cached text in	Audio in	Cached audio in	Text out	Audio out	Per session minute
`gpt-realtime-2.1`	$4.00	$0.40	$32.00	$0.40	$24.00	$64.00	—
`gpt-realtime-2.1-mini`	$0.60	$0.06	$10.00	$0.30	$2.40	$20.00	—
`gemini-2.5-flash-native-audio-preview-12-2025`	$0.50	—	$3.00	—	$2.00	$12.00	—
`grok-voice-think-fast-2.0`	$0	$0	$0	$0	$0	$0	$0.08
`grok-voice-think-fast-1.0`	$0	$0	$0	$0	$0	$0	$0.05

xAI bills speech-to-speech by the minute of audio rather than by token, which is why its token rates are a real zero rather than a missing figure.

Two provider charges on this pipeline are not in Talqing’s cost figures. The OpenAI entries run a separate transcription model so the call has a transcript at all; OpenAI bills it and reports no quantity we can meter. xAI charges per client text event, which nothing on the wire counts. Both land on your provider bill, not your Talqing balance.

The two grok-voice-* entries silently drop an image. They accept it, echo the item back with the picture removed, and then answer confidently about a photo they never saw. An agent on one of them cannot be sent an image at all.

Avatar

Provider	Model	Channel	Price
anam	`anam`	`video`	$0.16 per minute

Billed on wall-clock minutes for the whole call, idle included, on top of whatever produces the agent’s audio. The faces are a separate gallery — see video avatars.

Noise cancellation

Voice and video only, off unless you turn it on, and it needs an ai-coustics key like any other provider. enhancement_level runs 0.0 to 1.0 and defaults to 0.8.

Provider	Model	What it does
aicoustics	`quail_vf_s`	Isolates the one person talking to the agent, suppressing other voices, echo and background noise.
aicoustics	`quail_vf_l`	The larger Voice Focus model — same job, significantly more CPU, no documented accuracy gain.
aicoustics	`quail_l`	Keeps every voice and removes only noise — a speakerphone, a room, a distant speaker.

No price is listed because there is nothing for Talqing to meter: ai-coustics bills your own account directly. See audio processing.

Thinking and provider tools

How long the model may think, and the tools its provider runs for it.

Model fallbacks

A second provider per stage, and every rule it has to satisfy.

Provider keys

Bring the keys the models above run on.

Get started

Agents

Tools

Knowledge

Agent teams

Agent tasks

Telephony

Web & chat

Integrations

Email outbound

Observability

Guides

Platform

Two pipelines

The fields on each slot

Choosing

The priority lane

Reference

Language models

Speech-to-text

Text-to-speech

Realtime (speech-to-speech)

Avatar

Noise cancellation

Next

Thinking and provider tools

Model fallbacks

Provider keys

​Two pipelines

​The fields on each slot

​Choosing

​The priority lane

​Reference

​Language models

​Speech-to-text

​Text-to-speech

​Realtime (speech-to-speech)

​Avatar

​Noise cancellation

​Next

Thinking and provider tools

Model fallbacks

Provider keys

Two pipelines

The fields on each slot

Choosing

The priority lane

Reference

Language models

Speech-to-text

Text-to-speech

Realtime (speech-to-speech)

Avatar

Noise cancellation

Next