The cheapest-looking number in realtime voice is currently Google’s $0.005 per minute for Gemini 3.8 Live audio input. That number is real, but it is not a complete voice-session price — and comparing it directly with OpenAI’s $0.05 per minute GPT-Live-1 rate would mix different billing layers.
The September 20 snapshot shows three distinct pricing models. Google publishes separate per-minute audio input and output rates for Gemini 3.8 Live. OpenAI’s GPT-Live-1 charges for a full-duplex voice orchestration layer while backend model and tool usage remain separate. OpenAI’s GPT-Realtime-2.1 family instead exposes token-based text and audio pricing. A useful cost comparison therefore starts by identifying what each meter includes, not by sorting the smallest headline number.
This article is a dated interpretation of the current AI API market and Serverless market data. Resource pages remain the current-state reference because model prices and packaging can change.
Gemini’s minute rate is two-sided
Gemini 3.8 Live is priced at $0.005 per minute of audio input and $0.018 per minute of audio output. If a session contains one minute of user speech and one minute of generated speech, those two published meters sum to $0.023 before considering any other billable service that an application may add.
That is already a more useful figure than quoting $0.005/min alone. Voice conversations are rarely input-only, and output speech is 3.6 times the published input minute rate. A workload with long assistant responses will therefore have a different effective cost from a terse command interface even when both run for the same wall-clock duration.
Google lists the same $0.005 input and $0.018 output minute rates for Gemini 3.8 Live Extended Thinking. The differentiator in the current snapshot is capability: the Extended Thinking variant adds configurable background reasoning and asynchronous tool execution for more complex multi-step voice workflows. The published minute rates alone do not establish whether that extra reasoning changes total application cost through other usage dimensions, so it would be unsafe to infer a flat all-in session price beyond the documented audio meters.
GPT-Live-1’s $0.05/min is a layer, not an all-in model bill
GPT-Live-1 is listed at $0.05 per minute and billed per second. It provides the full-duplex interaction layer: simultaneous listening and speaking, interruption handling, transcripts, turn detection, WebRTC/WebSocket connectivity and telephony integration.
The critical pricing qualifier is that backend model and tool usage are billed separately. Multiplying $0.05 by call duration therefore calculates only the GPT-Live-1 layer, not the complete cost of an agent that also reasons with another model, invokes tools or uses additional services.
For a 10-minute call, the front-end layer alone is $0.50. That is a valid calculation from the published rate. Calling $0.50 the total OpenAI agent cost would not be valid without the backend workload. This distinction matters especially when comparing GPT-Live-1 with Gemini Live, where the snapshot records audio input/output minute meters as part of the model’s Live API pricing.
OpenAI also exposes narrower speech services. GPT-Live-Transcribe costs $0.017 per minute of realtime audio for streaming speech-to-text. That can be economical when the product needs transcription rather than a bidirectional voice agent, but it is not a substitute price for generated speech or reasoning.
OpenAI’s Realtime models use another cost language
The newer GPT-Realtime-2.1 family is harder to reduce to a single minute figure because the canonical pricing is token-based.
GPT-Realtime-2.1 Mini lists text at $0.60 per 1M input tokens, $0.06 per 1M cached input tokens and $2.40 per 1M output tokens. Audio is $10 per 1M input audio tokens and $20 per 1M output audio tokens.
GPT-Realtime-2.1 raises those rates to $4 per 1M text input tokens, $0.40 cached input, $24 text output, $32 audio input and $64 audio output. Both resources have a 128K-token context window in the current snapshot.
Those rates make one conclusion safe and another unsafe. It is safe to say Mini has materially lower published token prices than the full model. It is not safe to convert either model into a universal dollars-per-minute number without an authoritative mapping from the session’s audio to billable audio tokens and without knowing the text, cache and tool usage of the application.
This is why a voice-agent spreadsheet should not force every provider into one $ / minute column. Doing so would create precision that the source pricing does not support.
Use a bill-of-materials model instead of one headline rate
A better estimator separates the voice system into billable components.
| Component | Current published meter | What the number can establish |
|---|---|---|
| Gemini 3.8 Live audio input | $0.005/min | Cost of documented input-audio duration |
| Gemini 3.8 Live audio output | $0.018/min | Cost of documented generated-audio duration |
| GPT-Live-1 | $0.05/min | Cost of the full-duplex voice layer; backend model/tools extra |
| GPT-Live-Transcribe | $0.017/min | Streaming transcription cost |
| GPT-Realtime-2.1 Mini audio | $10/M input, $20/M output audio tokens | Token-metered realtime audio model cost |
| GPT-Realtime-2.1 audio | $32/M input, $64/M output audio tokens | Higher-priced token-metered realtime audio model cost |
For Gemini Live, an estimator can start with actual user-speaking minutes and generated-speaking minutes. For GPT-Live-1, it should start with connected voice-layer duration and then add the selected backend model and tool bill. For GPT-Realtime, it should use observed audio/text token consumption from representative sessions rather than inventing a minute conversion.
This approach also handles product design changes better. If an agent becomes more concise, Gemini output minutes may fall. If an OpenAI implementation changes its backend model, GPT-Live-1’s front-end minute rate may remain unchanged while the all-in bill changes. If a Realtime application improves caching or alters prompt/context behavior, token costs can move without any change in call duration.
The practical comparison is cost per completed task
For a prototype, published rates are enough to eliminate obviously mismatched architectures. For production, the useful denominator is usually cost per completed support case, booking, qualification call, translation session or other business task.
A representative test set should record user-audio duration, assistant-audio duration, connected session time, text/audio tokens where exposed, tool calls and successful task completion. That turns provider billing into an application-level metric instead of assuming that one minute of a quiet support call equals one minute of an interruption-heavy agent workflow.
The current market data suggests a sensible shortlist rather than a universal winner. Gemini 3.8 Live exposes unusually legible two-sided minute pricing. GPT-Live-1 packages a broader voice interaction layer but explicitly leaves backend usage outside its $0.05/min meter. GPT-Realtime-2.1 Mini offers substantially lower token rates than GPT-Realtime-2.1 when the application can use the smaller-cost model. GPT-Live-Transcribe is a separate low-cost path when speech-to-text is the actual requirement.
The decision rule is simple: compare like billing layers, measure a representative conversation, and optimize the cost of a completed task. A single headline $ / minute ranking would hide more than it reveals.