We use the Gemini Developer Live API, standard gemini-3.8-live, Python google-genai==2.19.0, AUDIO output and Achernar. We need to pass through the complete provider cost per call using prepaid customer credits. Please identify the authoritative billable meters and supported spending bounds for this exact model and endpoint. This is not a request for a downloadable invoice.
Synthetic reproduction
A standalone capture connects with a short synthetic system instruction, max_output_tokens=256, session resumption and compression configured, then sends “Hello.”, “こんにちは。” and “Thank you.” as separate inputs. No learner history or private material is used. We repeat with transcription configuration omitted, input only, output only and both.
First raw wire receipt from the transcription-omitted run:
{
"promptTokenCount": 522,
"responseTokenCount": 42,
"totalTokenCount": 564,
"promptTokensDetails": [
{"modality": "TEXT", "tokenCount": 332},
{"modality": "AUDIO", "tokenCount": 165}
],
"responseTokensDetails": [{"modality": "AUDIO", "tokenCount": 42}],
"thoughtsTokenCount": 136
}
Raw wire and parsed SDK metadata match. The prompt has 25 tokens not covered by its modality details. Across the three turns the gaps are 25, 39 and 51 in each of the four transcription configurations. Output transcription still arrives with its configuration omitted: that run produced 41 transcript characters, but no output TEXT token detail. Separate tool captures also include response totals without a response modality breakdown.
For a long-output request, a 256 output setting yielded 56 response tokens and 187 thinking tokens in one captured result. A 32 setting timed out without a usage receipt. Neither observation establishes a general spending ceiling.
Questions requiring a model-specific answer
- Are the 25/39/51 omitted prompt tokens billable? If so, which meter and price apply? If not, is the modality breakdown the complete billable input even when smaller than
promptTokenCount? - Where are exact input/output transcription token counts reported? The billing guide prices both as output text. Character counts cannot recover exact token counts. Does omitting the configuration prevent billing when transcript events still arrive?
- Are
thoughtsTokenCountandtoolUsePromptTokenCountadditional to the modality details, and under which rates? The Live reference defines total as prompt plus response; the example above excludes the separate thoughts count. How should a tool-only response without output modality details be charged without double counting? - What identifies the final per-generation billable receipt, including interrupted turns, tool continuations, delayed standalone usage, compression and session resumption? Is there a supported lookup for missing final usage?
- What provider-enforced bound covers all chargeable input, thinking, audio, transcription and tool continuations before work begins? Does
maxOutputTokenscap their combined generation cost on the standard Live endpoint? Does manual activity detection prevent generation before an explicit authorized boundary? Compression is not itself a per-request cap.
Please provide the correct calculation for the receipt above, including separately metered transcription, and a supported configuration that bounds each authorized unit of work. If either capability is unavailable, please confirm that limitation explicitly.
The observations come from synthetic captures made October 12, 2026. No keys, account IDs, learner transcripts, audio recordings or resumption handles are included here. Further synthetic metadata and reproduction code can be provided if useful.
References: Live usage fields, Live billing, standard model, pricing.