The August Voice Stack: Rebuilding Inbound Call Handling

Rebuilding inbound call handling on an AI voice stack got both easier and more confusing in the first week of August 2026. Three things converged: OpenAI’s full-duplex GPT-Live, which launched on July 8, the Grok Voice alias flip to Think Fast 2.0 executed on August 5, and a pair of OpenAI transcription models settling onto public price lists. None of the three is a product you can buy whole. What you buy is layers.

That distinction is the whole problem. A demo of a duplex voice model answering a call looks like a finished product, so buyers price it like one. In practice an inbound lead-handling system is a conversation layer, a transcription layer, and a write-back path into the CRM — three separately-priced things, two of which have list rates and one of which almost never does. Get the architecture wrong and you end up paying an orchestration fee for a system whose actual cost sits somewhere else entirely.

This guide covers what landed in the week to August 7, 2026, how the three layers fit together, what is genuinely purchasable today versus what is a signup form, a priced reference table across the whole stack with the per-1,000-minute arithmetic worked out, and a build-versus-platform decision framework using Vapi and Retell AI’s current published rates. Every price below is labelled by the surface it was read from.

The week to August 7 did not produce a single headline release. It produced three adjustments that only matter together, and each one changes a different layer of the same stack.

The first is architectural and slightly older: OpenAI’s GPT-Live launched on July 8, 2026 and, at the time of writing, is still the newest full-duplex voice system on the board. The second is a quiet substitution: xAI’s grok-voice-latest alias flipped to Think Fast 2.0 on August 5, so anyone pointing at the alias woke up on a new model. The third is pricing hygiene — the two transcription models OpenAI announced in late July now sit on the public API pricing page with per-minute rates, and gpt-live-transcribe appeared on a reseller surface on August 5.

Read individually, none of these is a reason to rebuild anything. Read together they define the shape of what is buildable in August 2026: the newest full-duplex conversational front-end is unavailable to developers, the duplex options a developer can actually call sit elsewhere — a different vendor or a parallel API SKU — and the transcription layer just got cheap enough that the interesting question is no longer whether to transcribe but when.

An inbound lead-handling system decomposes cleanly into three layers, and the decomposition is worth being pedantic about because vendors sell across the boundaries.

Layer one is the conversation. Something has to hold a natural spoken exchange with a caller — handle interruptions, know when to wait, know when to speak. Until 2026 this was almost always a cascade: a speech-to-text model, then a language model, then a text-to-speech model, chained in series. OpenAI’s own launch post describes that prior architecture plainly, and the description is the clearest statement of why duplex is a different thing rather than a faster version of the same thing.

A full-duplex system collapses that chain. GPT-Live listens and speaks at the same time and makes an interaction decision — speak, listen, pause, interrupt, call a tool — many times per second. It produces backchannels of the kind humans make while listening, and it waits rather than cutting in mid-thought. The architectural detail that matters most for a stack diagram is that it does not do the thinking: reasoning, search and agentic work are delegated to a frontier model running in the background, which at launch is GPT-5.5 — Instant tier on GPT-5.5 Instant, Medium and High on GPT-5.5 Thinking at medium and high effort. GPT-Live is a conversational front-end, not the brain.

Layer two is transcription. Even with a duplex conversation layer you still want text, and you want it for different reasons at different times. Mid-call text drives live routing, qualification logic, and anything that has to act before the caller hangs up. Post-call text drives the record, the summary, the search index and the compliance trail. These are different products with a roughly 3.8× price difference between them, which we work through in section 04.

Layer three is the write-back. This is the layer with no list price, because nobody sells it as a SKU. Everything that happens after the call ends — the record created or matched, the fields populated, the status advanced, the follow-up queued — is integration work against your CRM’s API. It is also, in our experience, where the majority of a voice project’s engineering hours actually go, and the reason so many voice pilots demo well and never reach production.

Start with the single most important buyer-facing fact, because it invalidates a lot of plans: as of this post’s publication date, GPT-Live has no public developer API. OpenAI describes it as coming soon and offers a notification signup form. No timeline has been published, and we are not going to guess one. The practical consequence is unambiguous — you cannot build GPT-Live into a custom inbound stack today. It ships inside ChatGPT Voice and nowhere else. Our customer-experience overview of GPT-Live covers what the model changes for the ChatGPT surface; this post is the build-side companion, and on the build side the answer today is that it is not an option. OpenAI has published a safety and system card for GPT-Live, which is worth reading if you are planning against a future API rather than building on one today.

What OpenAI does sell is a separate Realtime API audio SKU, gpt-realtime-2.1, along with a mini variant. Both are available now and both are token-metered rather than minute-metered. On the vendor pricing page read at the time of writing, the full model lists at $32.00 per million audio input tokens and $64.00 per million audio output tokens, with cached input at $0.40. The mini lists at $10.00 and $20.00 with cached input at $0.30 — 3.2× cheaper on both the input and output audio rates.

That cached-input rate deserves its own sentence. At $0.40 against $32.00, cached audio input on gpt-realtime-2.1 is an 80× discount on the uncached rate; on the mini it is roughly 33×. For an inbound agent that replays the same long system prompt, script and knowledge context on every single call, prompt-cache discipline is the single largest lever on the bill — larger than the choice between the full model and the mini for most call profiles.

The other genuinely purchasable duplex option is Grok Voice. On August 5, 2026 xAI flipped the grok-voice-latest alias to Think Fast 2.0, per xAI’s announcement. The published figures for 2.0: 82.9 on the Artificial Analysis Speech-to-Speech Quality Index against 75.7 for 1.0, time-to-first-audio down from 1.25 seconds to 0.70 seconds, and around 60% fewer reasoning tokens consumed. The raw model API lists at $0.08 per minute. Our deep-dive on Think Fast 2.0 and speech-to-speech agents works through the model itself; what matters here is that it is a duplex option a developer can actually call.

The latency improvement is the part worth taking seriously. Cutting time-to-first-audio from 1.25 to 0.70 seconds is a 44% reduction, and it lands in exactly the range where a caller stops experiencing a pause as a system delay. Quality-index points are an abstraction; half a second of silence at the top of a sentence is not.

The bar chart is deliberately awkward, and the awkwardness is the point. xAI’s own comparison chart omits Qwen Audio 3.0 Realtime Plus, which scores higher on the same index at 84.1 — but takes around four seconds to start speaking, roughly 5.7× longer than Grok Voice 2.0. That is not a clean win for either model. It is a straightforward accuracy-versus-speed trade, and for inbound calls specifically the trade resolves toward speed: a caller will forgive a slightly less nuanced answer long before they forgive four seconds of dead air.

One claim in the Think Fast 2.0 announcement needs an explicit caveat. xAI states that Grok Voice 2.0 is 1.5 to 2.0 times more accurate — that is, lower word error rate — than Deepgram Nova 3 and ElevenLabs Scribe v2, widening to roughly 10× in noisy conditions. Those figures come from xAI’s own unpublished harness. No absolute word error rates are published, no methodology is available, and we found no independent audit. Treat the direction as a vendor claim and the magnitude as unverifiable. If transcription accuracy is load-bearing for your use case, run your own audio through the candidates rather than trusting anyone’s chart, including that one.

OpenAI announced two transcription models together in late July 2026, and the pairing is a trap for anyone skim-reading a pricing page. The announcement thread introduces them as “two new transcription models in the API” — gpt-live-transcribe and gpt-transcribe — and the names are similar enough that they get used interchangeably. They should not be.

gpt-live-transcribe is low-latency streaming transcription, built for live microphone and call audio. gpt-transcribe is asynchronous batch transcription for completed audio files. On OpenAI’s API pricing page, read at the time of writing, the streaming model lists at $0.017 per minute and the batch model at $0.0045 per minute. The batch figure also appears on the OpenRouter listing for the same model at the identical rate — an unusually clean agreement between a vendor list and a reseller surface, and one we treat as corroborating rather than contradicting. Our comparison of GPT Transcribe against Whisper, Deepgram and Scribe v2 takes the accuracy side of that model apart in detail.

The architectural implication is that most inbound stacks should run both, not one. Stream only what the call logic actually needs to act on — the qualification questions, the routing signals, the intent markers — and batch the full recording afterwards for the record and the summary. A stack that streams every second of every call is paying a 3.8× premium on audio nobody needed in real time. A stack that batches everything cannot route intelligently while the caller is still there. The split is the design.

Both models are marketed as improved on accents, additional languages, short phrases and numbers relative to the prior Whisper-and-GPT-4o generation. That is a vendor claim; we located no independent word-error-rate benchmark for the new pair. Numbers in particular are worth your own testing, because inbound lead capture lives or dies on getting a phone number and a postal code right, and a model that is generally excellent and specifically bad at digits will fail in exactly the place that matters.

The table below is the spine of this post. Every published price is read from a vendor pricing page or launch announcement at the time of writing; every per-1,000-minute figure is our own arithmetic on those published rates, shown so you can check it. Token-metered models have no honest per-minute conversion without assuming a token rate, so we have not invented one.

Two things jump out of the table once the arithmetic is on the page. The first is that the cheapest complete duplex-plus-transcription combination a developer can assemble today runs on a vendor other than the one with the newest architecture — Grok Voice for the conversation at $80.00 per 1,000 minutes plus OpenAI batch transcription at $4.50, and the transcription is a rounding error against the conversation. The second is that the orchestration platform, at $50.00 per 1,000 minutes on Vapi’s published fee, costs more than the entire transcription layer several times over while doing none of the model work.

Every figure in the table above describes a layer with a rate card. The third layer does not have one, and that absence is not an oversight — it is because the work is specific to your CRM, your field schema and your pipeline rules, so no vendor can quote it.

The write-back layer is where a call becomes a lead. It has to decide whether this caller matches an existing record or needs a new one, extract the structured fields the sales process actually depends on, write them into the right objects, advance a status without violating whatever ordering rules the pipeline enforces, and queue the follow-up. None of that is voice work. All of it is the reason the voice work is worth paying for. We have written the build steps up separately in our guide to turning AI call notes into a CRM sync agent, and there is no point restating them here.

What is worth stating here is the sequencing consequence. Because the write-back layer has no list price, it tends to get scoped last and estimated optimistically, which is exactly backwards. The conversation and transcription layers are now commodity purchases with published rates and a handful of credible vendors — you can price them in an afternoon and swap them in a sprint. The write-back layer is bespoke, it is where the failure modes live, and it is the part that does not become cheaper when a new model ships. Scope it first and let it constrain the rest.

Once the layers are clear, the build-versus-platform question narrows usefully. You are not deciding whether to build a voice model — nobody is building a voice model. You are deciding whether to build the orchestration that sits between telephony, the models and your CRM, or to rent it from a platform and keep your engineering on the write-back layer.

Two platforms anchor the current pricing. On Vapi’s pricing page the hosting and orchestration fee is $0.05 per minute for voice calls, with SMS and chat at $0.005 per message; model-provider costs for STT, LLM and TTS are described as passed through at cost with no markup when you bring your own API key. Compliance sits outside that: HIPAA is $2,000 per month and Zero Data Retention $1,000 per month, which is $3,000 per month or $36,000 a year if you need both. The Build plan includes 10 concurrent lines, with additional lines at $10 per line per month.

On Retell AI’s pricing page pay-as-you-go voice agents run $0.07 to $0.31 per minute, and the page is explicit that the range spans model and voice choice rather than hidden fees. Twenty concurrent calls are included, additional concurrency is $8.00 per concurrency per month, and the enterprise tier is described as having no cap on concurrent calls. SMS is $0.01 per message. Phone numbers are $2.00 per month standard or $10.00 per month verified with a one-time $10.00 fee — which makes a verified number $130.00 in year one against $24.00 for a standard one. Chat-agent pricing varies by underlying model: GPT 5.5 at $0.052 per message, GPT 4.1 at $0.015, Claude 4.6 Sonnet at $0.03, Claude 4.5 Haiku at $0.007 and Gemini 2.5 Flash Lite at $0.006.

The concurrency arithmetic is the sharpest divergence between the two. To run 30 concurrent lines you buy 20 extra lines from Vapi at $10 each, or 10 extra concurrencies from Retell at $8 each — $200 per month versus $80, a $120 monthly and $1,440 annual difference before a single minute of audio is billed. On chat, the model choice swings harder than the platform choice: 10,000 chat messages costs $520.00 on GPT 5.5 against $70.00 on Claude 4.5 Haiku and $60.00 on Gemini 2.5 Flash Lite, a spread of roughly 8.7× decided by a dropdown. Our platform comparison of ElevenLabs, Vapi, Retell and Bland goes through the feature matrix properly; the numbers above are the pricing surface as published at the time of writing.

The most common budgeting error we see on voice projects is treating the platform’s headline per-minute number as the cost of a minute. It is not. On Vapi it is explicitly an orchestration fee with model spend passed through separately. On Retell the published range does fold model and voice choice in, which is why its band is so wide. Either way, the underlying STT, LLM, TTS and telephony spend is what moves.

Set against those components, the all-in figures circulating in third-party cost breakdowns are worth reading carefully. Several 2026 write-ups converge on roughly $0.07 to $0.25 and above per minute all-in for Vapi, and roughly $0.13 to $0.31 per minute all-in for Retell, once STT, LLM, TTS and telephony are added. Those are secondary aggregator estimates, not vendor-stated figures, and we cite them for the shape of the cost stack rather than as precise numbers.

Take the shape seriously even if you discount the digits. Subtract Vapi’s published $0.05 orchestration fee from that range and the model-plus-telephony residual is somewhere between $0.02 and $0.20 or more per minute — a tenfold spread on the part of the bill you control through architecture rather than vendor choice. For Retell, the aggregator range of $0.13 to $0.31 sits inside the upper three quarters of the vendor’s own published $0.07 to $0.31 band, which is a reasonable sanity check on both.

There is one unresolved pricing question we are not going to paper over. Grok Agent Builder advertised a $0.05 per minute platform price at its July 1 launch, set against the then-current model generation. With the default now flipped to Think Fast 2.0 at $0.08 per minute raw, we found no source confirming whether that platform price has changed. Treat it as unresolved and check the live page before you build a forecast on it.

Reading the week forward: the direction of travel is that the conversation layer is consolidating into a small number of minute-priced or token-priced commodity endpoints, while the capability gap between what ships in a consumer chat app and what ships in an API keeps widening. GPT-Live is the clearest example — the newest duplex experience available to a consumer in August 2026 is simply not available to a builder, at any price. If that gap persists, the practical consequence for anyone rebuilding inbound call handling is that the stack you design should assume model substitution rather than model selection. Keep the conversation layer behind an interface you can swap, keep the transcription decision independent of it, and put the durable engineering into the write-back layer, which is the only part that is still yours in twelve months.