when calculating total token used, should we include the summation of all the tokens - cached token for billing purposes

Thanks for reaching out, Utkarsh. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

AgentSessionUsage(model_usage=[LLMModelUsage(type='llm_usage', provider='chatgpt-us2.openai.azure.com', model='gpt-realtime', input_tokens=11293, input_cached_tokens=10240, input_audio_tokens=1143, input_cached_audio_tokens=704, input_text_tokens=10150, input_cached_text_tokens=9536, output_tokens=488, output_audio_tokens=328, output_text_tokens=160)]) for the above payload, what should be the calculation?

Thanks — I’m responding to your payload and the calculation question. You provided: • input_tokens = 11,293 • input_cached_tokens = 10,240 • input_audio_tokens = 1,143 • input_cached_audio_tokens = 704 • input_text_tokens = 10,150 • input_cached_text_tokens = 9,536 • output_tokens = 488 • output_audio_tokens = 328 • output_text_tokens = 160 Clarification: billing counts uncached input tokens, cached input tokens (billed at a lower cached rate), and output tokens. Audio/text breakdowns are subcomponents of the input/output totals and are not billed separately unless your pricing table lists separate audio/text rates—typically you use the totals (

input_tokens

,

input_cached_tokens

,

output_tokens

) for billing. I can compute a final cost — please tell me which per-1M rates to use for: • Input (uncached) rate • Cached input rate • Output rate If you’d like, I can use example rates and show the full numeric total. Which do you prefer? Sources: https://livekit.com/pricing.md| | Build voice, video, and physical AI | LiveKit, with interactive elements

get me an example

Thanks — I’ll use example per-1M rates and show the full calculation for your payload. Assumed example rates: • Input (uncached) rate: $5.00 per 1M tokens • Cached input rate: $0.50 per 1M tokens • Output rate: $2.50 per 1M tokens Values from your payload: • input_tokens = 11,293 • input_cached_tokens = 10,240 • output_tokens = 488 Compute uncached input: • Uncached input tokens = input_tokens − input_cached_tokens = 11,293 − 10,240 = 1,053 Compute costs: • Uncached input cost = (1,053 / 1,000,000) × $5.00 = $0.005265 • Cached input cost = (10,240 / 1,000,000) × $0.50 = $0.00512 • Output cost = (488 / 1,000,000) × $2.50 = $0.00122 Total cost: • Total = $0.005265 + $0.00512 + $0.00122 = $0.011605 Summary: • Uncached input tokens: 1,053 → $0.005265 • Cached input tokens: 10,240 → $0.00512 • Output tokens: 488 → $0.00122 • Total example cost: $0.011605 If you want this using your actual pricing rates, share them and I’ll recalc. , with interactive elements

how to get the token consumed for a call?

You're asking how to get the tokens consumed for a single call — here's how to do that. • Use the session's metrics event: listen for the

metrics_collected

(or provider-specific metrics) event on your AgentSession; the metrics payload includes

input_tokens

,

input_cached_tokens

, and

output_tokens

. • For OpenAI-style providers, the model response also returns

prompt_tokens

(input),

completion_tokens

(output), and `total_tokens`; LiveKit passes these through. • You can aggregate or report per-call usage with the

UsageCollector

utility, which processes

metrics_collected

events and sums token counts per call/service. Example (Python-style):

@session.on("metrics_collected")
def on_metrics(metrics):
    print("input_tokens:", metrics.input_tokens)
    print("input_cached_tokens:", metrics.input_cached_tokens)
    print("output_tokens:", metrics.output_tokens)

Notes: • Use

input_tokens

and

input_cached_tokens

together to compute billed input (uncached = `input_tokens - input_cached_tokens`; cached billed separately). • Audio/text subfields (e.g.,

input_audio_tokens

,

input_text_tokens

) are components of the totals; use the totals for billing unless you have separate audio/text rates. If you want, tell me your SDK/language and I’ll give a short code example for that environment. Sources: Realtime model metrics | LiveKit Documentation | OpenAI LLM models | LiveKit Documentation | Agents (Python) | livekit/livekit_composite | DeepWiki, with interactive elements

what about session_usage_updated?

You're asking about

session_usage_updated

— here's what it is and how to use it. • Purpose:

session_usage_updated

notifies you of aggregated usage changes for an AgentSession (e.g., updated totals for input tokens, cached input tokens, output tokens, and potentially audio/text subfields). It’s useful for tracking session-level billing and showing live usage to users. • How it differs from `metrics_collected`: ◦

metrics_collected

is emitted per LLM call or metric collection interval and contains granular metrics for that event. ◦

session_usage_updated

provides the updated session-wide totals (cumulative) after those granular metrics are applied. • What to read from it: ◦ Use the session totals (

input_tokens

,

input_cached_tokens

,

output_tokens

) from

session_usage_updated

for session-level billing and reporting. ◦ To compute per-call billed amounts, subtract the previous session totals from the new totals to get per-call delta, then apply your pricing (uncached = input − cached; cached billed at cached rate; output billed at output rate). • Example flow: a. Listen for

session_usage_updated

. b. On each event, compute delta = new_totals − previous_totals. c. Apply pricing to the delta tokens (uncached vs cached vs output). d. Update UI or billing records with the computed delta cost and new cumulative totals. If you want, tell me which SDK/language you use and I’ll give a brief code snippet showing how to listen for

session_usage_updated

and compute the per-call delta. , with interactive elements