Skip to content
TriedDocs
Reliability and latency

Estimated latency and cost

How the agent builder computes the Cost, Latency and Tokens estimates in its header, with the exact tables it uses.

The agent builder shows three live estimates in the Agent Details strip under the agent name: Cost per minute, Latency range and prompt Tokens. They update as you change the model, voice, transcription mode or prompt, before you place a single call. Hover each value for its breakdown.

The Estimated Latency popover in the agent builder: TTS, LLM and transcription

The same numbers come from the shared function estimateAgent in packages/shared/src/pricing.ts, and from the API at GET /agent-estimates/:agent_id (optional version).

Note

These are estimates from fixed tables, not measurements. Real calls also include network time, the end-of-turn wait and the phone network. Use Actual latency to see what calls measured.

How latency is estimated#

min = TTS + LLM.min + STT.min
max = TTS + LLM.max + STT.max

TTS (time to first audio)#

Voice model ms
sonic-turbo 120
sonic-3 150
eleven_flash_v2_5 (Standard) 180
eleven_turbo_v2_5, eleven_v4_turbo (Premium) 190
aura-2, inworld-tts-1.5-max 200
speech-2.8-turbo, inworld-tts-2 250
gpt-4o-mini-tts, s2.1-pro 300
tts-1 350
eleven_multilingual_v2, speech-2.8-hd 400

Unknown voice models count as 180 ms.

LLM (time to first token)#

Model ms
gpt-5.6-luna 300-500
gemini-2.5-flash 300-550
gpt-5.4-mini, claude-haiku-5-5, gemini-3.5-flash 350-600
gpt-4.1-mini 350-550
gemini-3-flash-preview 350-650
claude-haiku-4-5 400-700
gpt-4.1 450-700
gpt-5.6-terra (default) 500-800
gpt-5.5 600-900
claude-sonnet-4-6 600-1000
claude-sonnet-5-5 700-1100
claude-opus-5-5 900-1500

Unknown models count as 500-800 ms.

Transcription (stt_mode)#

Mode ms
fast (default) 130-300
accurate 300-500
custom 200-450

Example#

Premium voice (190) + gpt-5.6-terra (500-800) + fast (130-300) gives 820-1290ms, the default agent. With a Standard voice (180) it is 810-1280ms.

Speech-to-speech#

With a speech-to-speech engine there are no separate STT and TTS stages. The estimate is the engine's own range: gpt-realtime 500-900 ms, gpt-realtime-mini 400-700 ms.

Adjustments shown in the popover#

Setting Effect on the popover
High priority (model_high_priority on) LLM range multiplied by 0.75
responsiveness below 1 An extra row, Response wait time, of +(1 - responsiveness) x 5,500 ms
Pause Before Speaking An extra row, Pause before speaking (1st turn), of +the delay

The extra rows are shown for information and are not added to the range.

How cost is estimated#

The Cost popover lists each line and the total:

Line $/min
Voice infra 0.08
LLM (model) per model, see Billing
Voice (tier) per voice tier, see Billing
Knowledge base 0.01, when the agent has a knowledge base
STT (provider) only for stt_mode: custom with a non-Deepgram provider: AssemblyAI 0.004, Soniox 0.003, Azure 0.025, OpenAI 0.009

With high priority on, the LLM line is shown 50% higher. With a speech-to-speech engine, the LLM and TTS lines are replaced by one line: gpt-realtime $0.45, gpt-realtime-mini $0.15.

Prices are AUD, excluding GST. The default agent (gpt-5.6-terra, Premium voice, no knowledge base) estimates $0.30/min, or $0.24/min with a Standard voice. There is no telephony line: your own carrier bills the phone line. Billed amounts can include further lines such as prompt length scaling or guardrails; see Billing.

How tokens are estimated#

min = ceil(prompt characters / 4)
max = min + 2,100

The minimum covers the system prompt, tools and knowledge base context. The maximum adds about 2.1k tokens of conversation history. Prompts over 4,000 tokens increase the billed LLM cost (the llm_prompt_length line), so the token estimate is also a cost signal.

API#

curl https://api.tried.com.au/agent-estimates/agent_1a2b3c4d5e6f7a8b9c0d1e2f \
  -H "Authorization: Bearer $TRIED_API_KEY"
{
  "cost_per_min": 0.3,
  "cost_breakdown": [
    { "product": "Voice infra", "per_min": 0.08 },
    { "product": "LLM (gpt-5.6-terra)", "per_min": 0.1 },
    { "product": "Voice (Premium)", "per_min": 0.12 }
  ],
  "latency_ms": { "min": 820, "max": 1290, "breakdown": { "tts": 190, "llm": [500, 800], "stt": [130, 300] } },
  "prompt_tokens": { "min": 76, "max": 2176 }
}

The API response does not apply the high priority adjustments; the builder applies them on screen.