Meeting Processing — Cost Report

What it costs to run one meeting through the transcription pipeline, measured against meeting 850 with live LLM calls on 2026-09-09.

Meeting 850 — online GP consultation (ja-JP, 5.0 min): patient with recurrent unilateral pulsatile headache since their early 20s, now increasing in frequency, with visual aura, nausea, photophobia/phonophobia, and sleep-deprivation / red-wine triggers. Doctor rules out red flags, does a brief video neuro exam, diagnoses migraine with aura, prescribes a triptan + antiemetic + prophylactic, and advises a headache diary with follow-up. A short, single-complaint case with a clean two-speaker diarization — a light-to-typical load for the pipeline.

Pipeline

Meeting#... (meeting.rb:279)
  └─ TranscriptionJob         extract audio → GCS upload → STT BatchRecognize
       └─ TranscriptionPollJob    poll STT → Gemini speaker ID
            └─ TranscriptCorrectionJob   Claude Opus transcript correction
                 └─ TokenizationJob        local tokenize + Gemini romaji fallback
                      └─ PHIDetectionJob     Gemini PHI detection

ThirdPartyDetectionJob is not in this chain — it runs a local RF-DETR model (bin/detect_third_party), so it costs CPU/GPU time only, no metered API.

Billable service calls

#	Call site	Service	Billing basis
1	`transcription_job.rb:96` `stt_audio.attach`	Google Cloud Storage	Storage + write op for the WAV (purged after poll)
2	`transcription_job.rb:62` → `Transcriber.submit_batch`	STT v2 BatchRecognize, `chirp_3` + diarization + word offsets	Per audio-minute
3	`transcription_poll_job.rb:59` → `SpeakerIdentification.call`	Vertex AI Gemini `gemini-3.1-flash-lite`	Per token
4	`transcript_correction_job.rb:11` → `TranscriptCorrection.find_issues`	Vertex AI Claude `claude-opus-4-6`	Per token
5	`tokenization_job.rb:19` → `SuikaTokenizer.to_romaji` → `gemini_to_romaji`	Vertex AI Gemini `gemini-3.1-flash-lite`	Per token, per uncached CJK token
6	`phi_detection_job.rb:12` → `PHIDetection.run`	Vertex AI Gemini `gemini-3.1-flash-lite`	Per token

Not billed: Transcriber.poll_batch (GetOperation polling — quota only), MediaUtils ffmpeg/ffprobe (local compute), SuikaTokenizer MeCab path (local).

Price assumptions

Exact rates for the model builds in the code (chirp_3, gemini-3.1-flash-lite, claude-opus-4-6) are not separately published. These are early-2026 GCP / Vertex list prices used as proxies — treat as ±30%.

Service	Input	Output
STT v2 batch, `chirp_3`	$0.016 / min	—
Gemini flash-lite	$0.10 / 1M tok	$0.40 / 1M tok
Claude Opus	$15.00 / 1M tok	$75.00 / 1M tok
GCS standard	$0.02 / GB-month + $0.005 / 1k writes	—

STT free tier (first 60 min/month) is ignored as marginal cost.

Meeting 850 — input size

Metric	Value
Case	Online GP consultation — migraine with aura (new/worsening recurrent headache)
Video length (`original_duration_seconds`)	300.0 s (5 min 0 s)
STT wall-clock (`stt_elapsed_seconds`)	180 s
Language	ja-JP
Transcript segments (turns)	54
Transcript words	1,972
Transcript text	3,537 chars
`formatted_transcript` (fed to PHI detection)	4,105 chars
Tokens stored (`meeting.tokens`)	2,107 (453 unique)
Correction issues found	9

Measured token usage — meeting 850

Step	Model	Input tok	Output tok
Speaker ID	Gemini flash-lite	360	19
Transcript correction	Claude Opus	11,243	1,428
PHI detection	Gemini flash-lite	2,626	40
Romaji fallback	Gemini flash-lite	51 (×1 call)	13

Only 1 of 453 unique tokens ("拍") hit the Gemini romaji fallback; MeCab handled the rest.

Cost — meeting 850 (5.0 min)

Step	Calculation	Cost
Transcript correction (Claude Opus)	11,243 × $15/M + 1,428 × $75/M	$0.27575
STT BatchRecognize	5.0 × $0.016	$0.08000
PHI detection	2,626 × $0.10/M + 40 × $0.40/M	$0.00028
Speaker ID	360 × $0.10/M + 19 × $0.40/M	$0.00004
Romaji fallback	51 × $0.10/M + 13 × $0.40/M	$0.00001
GCS WAV (~9.6 MB, minutes)	storage + 1 write	~$0.00010
Total		≈ $0.3562

≈ $0.356 per meeting — $0.071 per audio-minute.

Cost split: transcript correction 77%, STT 22%, everything else < 0.1%.

Extrapolation by duration

STT scales exactly with duration; correction input scales ~linearly with transcript length; correction output grows sub-linearly. "Other LLM" is the Gemini speaker-ID + PHI + romaji calls combined.

Duration	STT	Correction (Opus)	Other LLM	Total	Per min
5 min (850)	$0.080	$0.276	~$0.0004	~$0.356	$0.071
15 min	$0.240	~$0.75	~$0.001	~$0.99	$0.066
30 min	$0.480	~$1.45	~$0.002	~$1.93	$0.064
60 min	$0.960	~$2.85	~$0.003	~$3.81	$0.064

Cost drivers & levers

Transcript correction dominates (77%). At Opus rates the single correction call costs $0.276 for a 5-min meeting — 3.4× the STT cost — and it scales ~linearly with transcript length, so the ratio only grows for longer meetings. Model choice on this one call is the biggest cost lever in the pipeline — a cheaper model here (Sonnet-class, or a Gemini flash/pro tier) would cut the meeting total by an order of magnitude.
STT is the second lever (22%). If the batch_recognize call qualifies for dynamic-batch pricing (~$0.003/min instead of $0.016/min), STT drops to $0.015 and the meeting to **$0.29**. Worth confirming which rate BatchRecognize + inline_response_config actually bills at.
Speaker ID, PHI, romaji are rounding error — together < $0.0004.
Retries multiply. Each LLM job has retry_on StandardError, attempts: 3. A transient failure on the correction job re-runs the whole Opus call (+$0.28 each). A worst-case double STT submission (documented SIGKILL race, transcription_job.rb:10-12) adds ~$0.08.
Prompt caching is unused. ClaudeClient.request_body sends no cache_control blocks (cache_creation_input_tokens: 0 observed). The ~600-token instruction header is re-billed at full input rate every call; caching it would shave only ~$0.009/meeting, so not worth it at current volume.

Reproduce

bin/rails runner '
m = Meeting.find(850)
segs = m.original_transcription

def g(p, model)
  x = GeminiClient.client.generate_content(model: GeminiClient.model_path(model),
    contents: [{role:"user",parts:[{text:p}]}],
    generation_config: {response_mime_type:"application/json", temperature:0}).usage_metadata
  {in: x.prompt_token_count, out: x.candidates_token_count}
end

def c(p, model)
  r = ClaudeClient.client.raw_predict(endpoint: ClaudeClient.model_path(model),
    http_body: {data: ClaudeClient.request_body(p), content_type: "application/json"})
  JSON.parse(r.data)["usage"].slice("input_tokens", "output_tokens")
end

puts g(SpeakerIdentification.build_prompt(segs, doctor_label: Doctor.name, patient_label: Patient.name), GeminiClient::DEFAULT_MODEL).inspect
puts c(TranscriptCorrection.build_prompt(segs), TranscriptCorrection::MODEL).inspect
puts g(PHIDetection.prompt(m.formatted_transcript), GeminiClient::DEFAULT_MODEL).inspect
'

bogdan/PRICING.md

Select an option

No results found