The short version: Google’s Gemini 3.8 Live bills $0.005 a minute for audio coming in and $0.018 a minute for audio going out. Both figures are published in Google’s own announcement and on its Gemini Live API pricing page, checked September 16, 2026. OpenAI’s competing GPT-Live-1 bills a flat $0.05 a minute of session time. The difference that matters is not the headline gap, it is what each meter counts. OpenAI charges for the clock. Google charges for the audio. So a minute in which nobody says anything, and real phone calls are full of them, costs five cents on one and half a cent on the other.
That is a ten to one gap on the quietest part of a conversation. It is also the reason the cost comparisons that appeared this week disagree with each other so badly. We found three widely repeated numbers for the same pair of products, and each one is measuring something different.
What Google shipped
On September 15 Google released three audio models: Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking, and Gemini 3.5 Transcribe. The announcement on Google’s blog describes native speech-to-speech models that can run a function in the background while still talking, handle 97 or more languages, and keep alphanumeric strings intact, which matters more than it sounds when a caller is reading out an account number or a postcode.
Google says the Extended Thinking variant ranks first on Artificial Analysis’ speech-to-speech leaderboard. The transcription model reports a 4.0 percent word error rate streaming and 2.6 percent on a finished file.
None of that is the interesting part for a business owner. The interesting part is the billing structure, because Google did not simply undercut OpenAI. It changed the unit.
What Gemini Live API pricing counts, and what OpenAI’s does not
Read the two pricing pages side by side and the shape of the difference is immediate.
OpenAI’s developer pricing page lists GPT-Live-1 at five cents a minute, billed per second, with the backend model and any tools charged separately. The meter starts when the session opens and stops when it closes. It does not care whether anyone is speaking. Silence, thinking time, a pause while your calendar is queried: all of it bills at the same five cents a minute.
Google’s page meters audio tokens instead. Input audio runs $0.005 a minute, or $3.00 per million tokens. Output audio runs $0.018 a minute, or $12.00 per million. The input side streams for the whole call because the microphone is open. The output side, which is the expensive one at roughly three and a half times the input rate, only accrues while the assistant is actually speaking.
One vendor sells you time on the line. The other sells you the sound.
So the honest comparison is not one number against another. It depends on how much of your call is talk.
Why the published comparisons disagree
Three figures have been circulating since the launch, and it is worth knowing which is which before you quote any of them.
The first is Google’s own chart, reported as $0.84 an hour for Gemini 3.8 Live against $5.83 an hour for OpenAI, measured per hour of input audio. Work backwards from Google’s published rates and $0.84 an hour is $0.014 a minute, which only balances if you assume the assistant speaks for half the session. That is a reasonable assumption. It is still an assumption, and it is not labelled on the chart.
The second is that $5.83 an hour figure for OpenAI, which works out to about $0.097 a minute. OpenAI’s own page says five cents. The gap is the backend model, because GPT-Live-1 is a voice layer that delegates the actual thinking elsewhere. Both numbers are correct. They are the same product with and without its brain attached, which is worth remembering whenever a voice price looks suspiciously round.
The third is a claim that Gemini is thirteen times cheaper, built by comparing it against OpenAI’s older Realtime models, which are still sold per audio token rather than per minute. Those models exist and are still listed, but they are not what OpenAI launched to compete here. Comparing a new model to a predecessor makes any gap look larger than the choice a buyer is actually facing.
The same forty calls, priced both ways
Two days ago we priced a month of missed calls for a small service business: forty calls averaging four minutes, 160 minutes of talk time, which came to eight dollars on GPT-Live-1 plus roughly a dollar of reasoning. Same baseline, run through Google’s meter, assuming the assistant speaks about 40 percent of the time, which is typical when the caller is supplying an address and a description of the problem.
Input runs for all 160 minutes at half a cent, which is 80 cents. Output runs for 64 of those minutes at 1.8 cents, which is $1.15. The month comes to about $1.95 against eight dollars.
Push it to the theoretical limit and the conclusion holds. If the assistant somehow spoke for every second of all 160 minutes, Google’s audio bill would be $3.68, still under half of OpenAI’s eight dollars. There is no talk ratio at which Gemini’s audio layer reaches OpenAI’s flat rate, because even continuous speech only gets you to $0.023 a minute.
The one line that could flip it
Neither headline covers the thinking, and that is where the answer could change.
Gemini 3.8 Live bills generated text at $4.50 per million tokens, inside the same meter. For that to push Google past OpenAI’s five cents a minute, the model would need to generate more than six thousand tokens of text per minute of conversation, and closer to eight thousand at a realistic talk ratio. Six thousand tokens is roughly 4,500 words. Per minute. For a booking or intake call, nothing comes close. For an agent doing heavy deliberation and long tool chains on every single turn, it is at least worth measuring rather than assuming.
That is the number to test if you are building. Log your token counts for a week and see where you actually land.
If you are buying rather than building
Most owners reading this will never touch either API, and the practical conclusion has not changed just because a price fell again. The model was already the cheapest part of an AI phone agent. It just got cheaper, which mostly widens the gap between what the technology costs and what a finished product costs, for reasons that remain good ones: a phone number, calendar integration, escalation logic, and someone to call when it books two jobs into the same slot.
What a cheaper meter does change is leverage. When the underlying cost of a minute drops by a factor of four, per-minute caps and overage charges in a vendor’s contract get harder to defend. It is a fair question to ask at renewal.
And the point of putting a machine on the after-hours line was never to take the phone away from whoever answers it during the day. Those calls are going to voicemail right now, and most people calling a service business in the evening do not leave one. They call the next number. Catching them, and handing your team the conversations that are actually worth a human being, is the whole job. A cheaper model makes that easier to justify. It does not change what good looks like.
For more on reading a 2026 pricing page, see our guide to what AI software is actually worth paying for and what happened when the cost of transcribing an hour of audio fell to ten cents.
Frequently Asked Questions
How much does Gemini 3.8 Live cost per minute?
Google publishes two rates rather than one: $0.005 per minute for audio input and $0.018 per minute for audio output, equivalent to $3.00 and $12.00 per million tokens respectively, checked on Google’s pricing page September 16, 2026. Because the output rate only accrues while the assistant is speaking, your actual per-minute cost depends on how much of the call is the model talking. At a typical 40 percent talk ratio that works out to roughly $0.012 a minute, and even continuous speech only reaches $0.023.
Is Gemini 3.8 Live actually cheaper than GPT-Live-1?
On the voice layer, yes, and by more than the headline suggests. OpenAI lists GPT-Live-1 at a flat $0.05 per minute of session time regardless of whether anyone is speaking, so a four minute call costs twenty cents there against roughly five cents on Gemini. The caveat is that neither price includes the reasoning: OpenAI bills a separate backend model entirely, while Google bills generated text at $4.50 per million tokens within the same meter, so an unusually chatty agent narrows the gap.
Why do the cost comparisons online disagree so much?
Because they are measuring three different things. Google’s own chart uses cost per hour of input audio, which quietly assumes the assistant speaks about half the time; the $5.83 an hour figure quoted for OpenAI includes a backend model that OpenAI’s own $0.05 a minute listing excludes; and the largest claimed gaps compare Gemini against OpenAI’s older Realtime models, which are billed per audio token rather than per minute. Always check whether a voice price includes the model doing the thinking, because that is usually about half the bill.
Should a small business switch voice providers over this?
Almost certainly not by itself, because if you bought a finished answering service you are not paying model rates anyway, and a few dollars a month of underlying cost is not what your invoice is made of. The useful move is to find out what your vendor charges for minutes or calls above your plan, since the wholesale cost of those minutes just dropped sharply and that is a reasonable thing to raise at renewal. If you are building your own agent, the switch is worth pricing properly, using your own logged token counts rather than anyone’s comparison chart.
We are curious about one thing. If you already run an AI agent on your phone line, do you actually know what a single call costs you, or does it just arrive as one line on a monthly invoice?
