Gemini 2.5 Flash Audio Cost
Sending audio to native Gemini 2.5 Flash ($1/M audio tokens), or transcribing first? This page computes both paths per request and tells you which wins for your audio length.
For teams processing audio with LLMs — compares transcribe-then-text against a native audio model, since native audio tokens are priced so differently per provider that the cheaper path isn't obvious.
Model prices from OpenRouter · updated 2026-07-13
01Your audio & models
02 Transcribe first, or send it raw?
Native Gemini 2.5 Flash wins by 66% — its audio tokens are cheap enough to beat transcribing.
At 100,000/mo: $3,108 transcribe vs $1,066 nativecheaper path saves $2,042/month
Related cost calculators
STT per-minute and native audio $/M-token rates hand-verified (checked 2026-07-13), re-audited quarterly: OpenAI pricing · OpenAI realtime audio tokens · Gemini pricing · Gemini token counting. Native models' text rates from OpenRouter, snapshot 2026-07-13. Audio tokens = minutes × 60 × tokens/sec (10 OpenAI, 32 Gemini). All math runs in your browser.
How the math works
Two ways to answer from 5 minutes of audio: transcribe it with GPT-4o Transcribe ($0.006/min) and feed the text to GPT-4.1 Mini, or send the audio straight to Gemini 2.5 Flash. Transcribe-then-text costs $0.03/request here; native Gemini 2.5 Flash costs $0.01. Native wins by 66%.
The catch is how native audio tokens are priced. Gemini 2.5 Flash bills audio input at $1/M and counts 32 tokens per second, so 5 minutes is 9,600 audio tokens = $0.009600 — versus just $0.03 to transcribe the same audio and $0.001080 to process the 900-token transcript. Gemini's audio tokens are cheap enough that native can undercut transcribe-then-text.
Same baseline, both decomposed: two-stage = STT ($0.03) + text model ($0.001080). Native = audio input ($0.009600) + prompt ($0.000060) + answer ($0.001000). The cheaper path is native Gemini 2.5 Flash at $1,066/month vs $3,108.
Cost isn't the whole story. Native audio models hear tone, emotion, overlapping speakers and non-speech cues that a transcript throws away, and they answer in one call instead of two (lower latency, no transcription errors compounding). If those matter for Gemini 2.5 Flash audio, native can be worth a premium — this page tells you what that premium is, not whether to pay it.
STT per-minute rates and native audio $/M-token rates are hand-verified against provider docs (checked 2026-07-13), re-audited quarterly; the native models' text in/out rates come from the daily-synced catalog (updated 2026-07-13). All math runs client-side with tested code.
Frequently asked questions
Is transcribe-then-text or native audio cheaper for Gemini 2.5 Flash audio?
Native Gemini 2.5 Flash is cheaper here — $0.01/request vs $0.03 for transcribe-then-text (66% less), because Gemini's audio tokens are cheap. Tune the sliders for your audio length and answer size.
Why are native audio tokens so much more expensive?
Audio carries far more than words — the model processes acoustic detail, so providers count and price audio tokens separately. Gemini 2.5 Flash charges $1/M audio input at 32 tokens/sec. A minute of audio is 1920 audio tokens; transcribing it instead yields only ~180 text tokens, which a text model processes for a fraction of the price.
When is native audio worth the premium?
When the acoustics matter: tone and sentiment analysis, speaker diarization, detecting laughter/hesitation/background sound, or when transcription errors would break the task. Native also answers in one round-trip instead of transcribe-then-generate, cutting latency. For plain "what did they say, now summarize it" tasks, transcribe-then-text is usually cheaper.
Does the transcription model choice matter?
Some — GPT-4o Transcribe is $0.006/min; the mini transcription model is cheaper still. But at these volumes the STT cost is often small next to the downstream text model, so the bigger lever is which path (and which native model) you pick.
Are these prices current?
STT per-minute and native audio $/M-token rates are hand-verified against provider docs (checked 2026-07-13) and re-audited quarterly. The native models' text rates sync daily from OpenRouter. All math runs client-side with tested code.