September 11, 2026 ·

openaivoicepricingapicustomer-serviceagents

GPT-Live-1 didn't just cut the price of voice — it changed the meter, and that decides which calls got cheaper

TL;DR: GPT-Live-1 reached general availability in the OpenAI API on 10 September 2026 at $0.05 per minute, billed per second. The capability jump is real: turn-taking latency 1.41s → 0.798s, Full Duplex Bench interactivity 45.4% → 80.10%, Tau3 voice 45.7% → 86.2%. But the change to model first is the unit. GPT-Realtime-2.1 billed audio tokens ($32/$64 per M), so dead air was nearly free and talking was expensive. GPT-Live-1 bills wall-clock, so silence and speech cost the same — hold queues, IVR waits and thinking pauses just became billable, and your silence ratio decides whether your bill went up or down. Second: $0.05 is the mouth, not the agent. Reasoning is delegated to a backend model on its own meter — the same ten-minute call is $0.60 paired with Luna and $1.40 paired with Astra. Third, the honest capability read: 32.0% on Tau Banking, best in class and still a two-thirds failure rate. For buyers: measure your silence ratio before you model anything, pair a cheap backend by default, and note that legacy realtime models retire 20 January 2027.

Two unbundlings on the same day

On 10 September OpenAI shipped the Agents API, which prices the harness at zero and meters the sandbox. The same day it shipped GPT-Live-1, which prices the voice layer at a flat rate and meters the reasoning separately.

The shape is identical. In both, OpenAI took a thing that used to be one bill and split it into a headline component and a variable one — and in both, the headline component is not the one that determines what you pay. Worth noticing as a pattern, because it is now twice in a day.

The meter changed, and that is the story

Here is what voice cost before.

GPT-Realtime-2.1 billed audio tokens: $32 per million input, $64 per million output (the mini ran $10/$20). Audio tokenizes at roughly one token per 100ms of user speech and one token per 50ms of assistant speech, so a minute of the customer talking is about 600 tokens and a minute of the agent talking is about 1,200 tokens. Add re-sent text context each turn, and measured production agents landed around $0.06–$0.11 per minute on the flagship.

Here is what it costs now: $0.05 per minute of session, billed per second. Not per token. Per elapsed second of the session being open.

Set the two meters against a single ten-minute support call:

Minute type	Old (token-metered)	New (wall-clock)
Customer speaking	~$0.019	$0.05
Agent speaking	~$0.077	$0.05
Silence — hold, thinking, reading an account number	~$0.00	$0.05

The third row is the one to take to a planning meeting. Under token billing, dead air was free, and a well-designed voice agent that let customers pause was economically rewarded for it. Under wall-clock billing, a four-minute pause in a ten-minute call costs $0.20 — the same as four minutes of dense conversation.

So the honest answer to “did voice get cheaper” is: it depends on a number you almost certainly are not measuring. Call it the silence ratio. High silence — IVR front-ends, hold queues, form-filling, anything with waits — got more expensive. Low silence — fast conversational triage, dense back-and-forth — got cheaper and predictable, which for anyone who has tried to forecast a token-metered voice bill is the larger of the two wins.

Flat wall-clock pricing also removes an ugly incentive. Under token billing, a terser agent was a cheaper agent, which quietly pushed teams toward clipped, unnatural responses. That pressure is now gone. Verbosity is free; time is the cost.

$0.05 buys a mouth, not an agent

The second correction to make before budgeting: GPT-Live-1 does not think. It listens and speaks, full-duplex, and delegates reasoning and tool use to a backend agent whose Responses calls bill at that model’s normal rates.

Ten-minute call, twenty backend turns at roughly 2,000 input and 500 output tokens each — 40,000 input, 10,000 output:

Paired backend	Backend cost	Voice cost	Total	Voice as % of bill
GPT-5.6 Luna ($1/$6)	~$0.10	$0.50	$0.60	83%
GPT-6 Astra ($10/$50)	~$0.90	$0.50	$1.40	36%

Same voice layer, 2.3x the total. The pairing decision is a bigger lever than the price cut everyone is quoting, and it points the same direction as the 272K threshold arithmetic on Astra: long sessions accumulate context, and accumulated context is where voice agents quietly become expensive.

The design that falls out is a two-tier backend — a cheap model handling routing, acknowledgement and lookup, escalating to a frontier model only when the turn genuinely requires reasoning. Most voice turns are not reasoning. They are “let me pull that up.”

What actually improved

The benchmark set is unusually coherent for a launch:

Metric	GPT-Realtime-2	GPT-Realtime-2.1	GPT-Live-1
Turn-taking latency	—	1.41s	0.798s
Full Duplex Bench interactivity	—	45.4%	80.10%
Tau3 voice (pass@1)	42.4%	45.7%	86.2%
Tau Banking voice, 97 tasks	10.3%	12.4%	32.0%
Conversational Dynamics (Artificial Analysis)	—	95.7%	97.3%

The latency number carries the most product weight. Crossing from 1.41 seconds to 0.798 takes the pause through the rough one-second boundary where a delay stops sounding like a system and starts sounding like a person thinking. That is not an incremental improvement; it is a category change in how the interaction reads.

The architectural reason it moved this far is the same reason the old approach plateaued. Chaining speech-to-text, then reasoning, then text-to-speech accumulates latency at every handoff and loses paralinguistic context — tone, hesitation, the start of an interruption — in the conversion to text. A full-duplex model listens and speaks in one pass, so barge-in and backchannelling are native behaviours rather than heuristics layered on a transcript boundary. It also means the transcription price floor is now a separate market from the voice-agent market: Gemini 3.5 Transcribe and the sub-$0.30/hour ASR tier compete for pipelines that still convert to text, which full-duplex agents increasingly do not.

The number that should slow a deployment down

Tau Banking: 32.0%.

Best in class. 2.6x the previous generation. And a two-thirds failure rate on banking tasks.

Hold it next to 86.2% on general Tau3 customer-service tasks and the profile is clear: this model is excellent at conversation and poor at domain procedure. The two scores describe different things, and only one of them has improved to the point of autonomy.

The deployment shape that follows is narrow. Let GPT-Live-1 own conversation, intent capture, triage and handoff. Let deterministic systems own anything that moves money, changes an account of record, or is subject to audit. Vendors building on this — Sierra and Decagon chief among them — will ship the upgrade fast because the latency improvement is immediately audible to their customers, and Decagon’s valuation has been priced on exactly this trajectory. Ask them for domain-task pass rates, not conversational ones. The gap between the two is where the incidents live.

Constraints to design around

Rate limits are concurrent sessions, not tokens: 25 / 50 / 200 / 300 / 500 across tiers 1–5. Five hundred simultaneous calls is the top self-serve ceiling; contact centres above mid-size need a negotiated arrangement before launch.
Single endpoint: only v1/live/sessions. It does not drop into existing Responses or Chat Completions plumbing.
Knowledge cutoff 31 July 2025. Anything current must arrive through the backend agent’s tools, not the voice model’s memory.
Audio and text only. No image, no video.
Migration clock: legacy audio and realtime models retire on 20 January 2027, part of the deprecation calendar we mapped when the Assistants API shut down. Roughly four months of overlap from GA — workable, but tight for regulated deployments that need re-validation.

What to do with this

Measure your silence ratio on real traffic first. It is the single number that determines whether the meter change helped or hurt you, and every other estimate depends on it.
Trim dead air, not words. Under the old meter you shortened responses. Under this one you shorten sessions — close and reopen across long waits rather than holding an open session through a hold queue.
Default the backend to cheap, escalate deliberately. A Luna-class model with an Astra escape hatch beats Astra everywhere, by roughly 2x on a typical call.
Benchmark on your own domain tasks. 86.2% general and 32.0% banking is the same model. A generic score will not tell you which one your workload resembles.
Plan capacity in concurrent sessions, and check the tier-5 ceiling against your peak before you commit.

For the surrounding market, our best AI audio tools guide covers the voice and transcription layers, the ElevenLabs review tracks the incumbent in synthesis, and the ChatGPT review tracks OpenAI’s model tiers and what they cost. For the other half of OpenAI’s 10 September unbundling, see the Agents API’s free harness and metered sandbox hours.

Frequently asked questions

Is $0.05 per minute actually cheaper than what I was paying for realtime?

For a busy call, yes. For a quiet one, no — and that reversal is the whole story. GPT-Realtime-2.1 billed audio tokens at $32 per million input and $64 per million output, with audio converted at roughly one token per 100ms of user speech and one token per 50ms of assistant speech. That means a minute of user talking is about 600 tokens and a minute of assistant talking is about 1,200 tokens. Measured production agents landed around $0.06 to $0.11 per minute on the flagship once context re-sends were counted, and roughly $0.02 to $0.05 on the mini with prompt caching working. Against that range, a flat $0.05 is a cut for conversation-dense workloads. But token billing charged you for audio that existed. Wall-clock billing charges you for time that elapsed. A ten-minute support call where the customer spends four minutes on hold, reading an account number, or thinking, used to cost near zero for those four minutes. It now costs $0.20. Whether your bill went up or down depends entirely on your silence ratio, and almost nobody currently measures that.

What does the $0.05 actually buy, and what does it not?

It buys the voice layer: listening and speaking, full-duplex, in a single model. It does not buy thinking. GPT-Live-1 delegates reasoning and tool use to a backend agent, and OpenAI's documentation is explicit that backend Responses calls are billed at the normal pricing for whichever model and tools you configure. So the real price of a voice agent is two meters added together, and the second one dominates the decision. Work a ten-minute call with twenty backend turns at roughly 2,000 input and 500 output tokens each — 40,000 input and 10,000 output. Paired with GPT-5.6 Luna at $1/$6 per million, that backend is about $0.10, so the call is $0.50 of voice plus $0.10 of reasoning: $0.60, and the voice layer is 83% of it. Paired with GPT-6 Astra at $10/$50, the backend is about $0.90 and the call is $1.40, where voice is 36%. Same voice bill, more than double the total. The model you pair is a bigger lever than the voice price you just celebrated.

How much better is it, really?

Substantially, and the latency figure is the one that changes user behaviour. Turn-taking latency drops from 1.41 seconds on GPT-Realtime-2.1 to 0.798 seconds — through the roughly one-second threshold where a pause stops reading as a system delay and starts reading as a person thinking. Full Duplex Bench interactivity goes from 45.4% to 80.10%, a roughly 30-percentage-point overall Full Duplex Bench improvement. Task performance moves further than the audio numbers: Tau3 voice pass@1 goes from 45.7% to 86.2%, and Artificial Analysis' Conversational Dynamics reads 97.3% against 95.7%. The architectural reason is worth understanding, because it is why the jump is this large. The old pattern chained speech-to-text, then a reasoning model, then text-to-speech, accumulating latency and losing context at each handoff. A full-duplex model listens and speaks in one pass, so barge-in, backchannels and interruption recovery are native rather than bolted on.

Where is it still not good enough to run unattended?

Tau Banking is the number to put in front of anyone planning autonomous financial voice support. Across 97 banking tasks GPT-Live-1 scores 32.0%. That is 2.6x GPT-Realtime-2.1's 12.4% and best in class, and it is still a two-thirds failure rate. Read the two benchmarks together: 86.2% on general customer-service tasks and 32.0% on banking-specific ones is a model that handles conversation extremely well and domain-specific procedure poorly. The deployment that follows from that is narrow and specific — let GPT-Live-1 own the conversation, triage, and handoff, and let deterministic systems own anything that moves money or changes an account of record. Voice agents that fail visibly at the transfer are survivable. Voice agents that fail confidently at a transaction are not.

What are the capacity limits, and how do I plan against them?

Rate limits are expressed in concurrent sessions rather than tokens per minute, which is a different planning unit than most teams have wired into their dashboards: 25 concurrent sessions at tier 1, 50 at tier 2, 200 at tier 3, 300 at tier 4 and 500 at tier 5. For a contact centre this is the constraint that actually binds. Five hundred simultaneous calls is the ceiling of the highest self-serve tier, so any operation above mid-size needs a negotiated arrangement or a second provider in the path before launch, not after. Two other limits are worth noting at design time: GPT-Live-1 supports only the v1/live/sessions endpoint, so it does not drop into existing Responses or Chat Completions plumbing, and the knowledge cutoff is 31 July 2025 — old enough that anything current has to arrive through the backend agent's tools rather than the voice model's memory. Modalities are audio and text only; image and video are unsupported.

Do I have to migrate, and by when?

Yes, and the clock is already published. OpenAI's deprecation calendar retires legacy audio and realtime models on 20 January 2027, which gives roughly four months of overlap from general availability. That is a workable window but not a generous one for anyone with a certified or regulated voice deployment, where re-validation takes most of a quarter. The sensible sequence is to instrument before you port. Measure your current silence ratio on real traffic first, because that single number tells you whether the meter change helps or hurts you, and it is the input every other estimate depends on. Then price the paired backend model deliberately rather than defaulting to the frontier tier — most voice turns are routing and acknowledgement, not reasoning, and a cheap backend with an escalation path to an expensive one is usually the right shape. Then run the migration against your own task set rather than Tau3, since the gap between 86.2% general and 32.0% banking is exactly the gap a generic benchmark hides.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.