OpenAI launched GPT-Live: full-duplex voice is on the Free plan. What changes for your Latam SaaS

July 30, 202616 minAI, OpenAI, GPT-Live, voice, SaaS, Latam, unit-economics

Short answer (60 seconds): On July 8, 2026 OpenAI launched GPT-Live, its family of full-duplex voice models (listening and speaking at the same time), with two versions: GPT-Live-1 as the default for Go/Plus/Pro and GPT-Live-1 mini as the default for the Free tier (Introducing GPT-Live). It is the first time OpenAI ships full-duplex to the Free tier (Advanced Voice Mode, the prior generation, was turn-based and never shipped to Free) — and, per Reuters, the first time a ChatGPT Voice capable of "listening to you while it talks" reaches a massive installed base. For a Latam SaaS founder that means the cost of prototyping a voice agent in Spanish/Portuguese drops to USD 0 in the demo phase and forces you to recompute production unit economics.

On July 30, 22 days after launch, I still see Latam founders who have not recomputed anything. The question I got three times this week was some variant of "do I stop paying Retell/Vapi and use GPT-Live directly?". The short answer: depends on the case. This post gives you the framework with real numbers, the comparison table against wrappers, and the operational checklist to decide this week.

Disclosure: I did not have access to GPT-Live-1 via the API at the time of writing — developer access is still a notification form and not an open endpoint (confirmed by Digital Applied). The API prices I cite are those of the gpt-realtime-2.x family (OpenAI's prior full-duplex generation, in production since May) taken from the OpenAI pricing page captured July 13 and Aireiter. Where I cite a number without an exact source, I flag it inline.

What happened this week (and why it matters)

OpenAI launched GPT-Live on July 8, 2026, opening two new voice models to the global ChatGPT base (Introducing GPT-Live, Reuters, Jul 8 2026, Business Insider: The Whirlwind 72 Hours of Rival AI Announcements). They positioned it as the replacement for Advanced Voice Mode, with the promise that "it feels like talking to a person" — because it can literally listen to you while it talks.

Four launch facts that matter for the Latam founder:

1. Two models at once, with Free already in. GPT-Live-1 becomes the default for paid plans (Go, Plus, Pro). GPT-Live-1 mini becomes the default for the Free plan. This is the first time OpenAI ships full-duplex at a $0 SKU (Advanced Voice Mode, the prior generation, was turn-based and never shipped to Free) — not a stripped demo, but the same model running on shared infrastructure.

2. Real full-duplex, not turn-based in disguise. GPT-Live makes "speak / keep listening / pause / interrupt / call a tool" decisions many times per second. Advanced Voice Mode (the prior generation) waited for the end of your turn. The practical difference: it lets you interrupt at any time, without the robotic "I had finished speaking" pause (MarkTechPost via Kie.ai).

3. Hands-off to GPT-5.5 for reasoning. When a question needs web search, long logic or a tool call, GPT-Live delegates to GPT-5.5 running in the background and returns to the conversation when the sub-agent finishes. The voice loop does not break. It is the architecture the rest of the vendors (Vapi, Retell, LiveKit) charge you to assemble as STT → LLM → TTS pipeline.

4. No API yet, but the queue is open. The official post says "coming soon to the API, sign up to be notified". Nobody outside OpenAI can build a customer-facing agent on GPT-Live today. For real production you stay on the gpt-realtime-2.x family, already on the API since May and keeping the same full-duplex architecture.

The macro context: launch week was, per Business Insider, "72 hours of non-stop AI announcements" — OpenAI with GPT-Live and GPT-5.6, Meta with Muse Spark 1.1, SpaceXAI with Grok 4.5, Anthropic with Reflect. If you are a Latam SaaS founder, the cost of building voice-first in Spanish/Portuguese is falling structurally — the question is no longer whether you can, it is which stack you should use.

GPT-Live vs Vapi vs Retell vs LiveKit — the honest table

Before deciding "build vs buy vs OpenAI direct", look at what each does. The table compares the options a Latam SaaS founder evaluates in July 2026:

Quick read of the table:

• GPT-Live Free is the cheapest demo in the history of voice-first SaaS. For a hackathon or prototype, you do not spend a cent. The problem shows up when you want a real call with your Twilio number, persistent transcription and analytics — there you still need a wrapper.
• Vapi charges "little" on platform ($0.05/min) but the real bill lands at $0.13–$0.32/min once you add Twilio + Deepgram STT + LLM + ElevenLabs TTS (Retell, caller.digital). Classic "headline price vs real cost".
• Retell gives it bundled at $0.07–$0.31/min by voice and LLM tier. The fastest path to production if you do not want to manage 4 providers (Dev.to, Apr 2026).
• LiveKit is the "build" option with the lowest cost ($0.05–$0.22/min all-in) but requires 2–4 weeks of engineering (Fora Soft, 2026).
• Latency matters more than price for Latam. Vapi/Retell average 500–900 ms. GPT-Live native drops to ~300 ms. In a sales agent, that difference is the line between "feels like a person" and "feels like a bot".

Table disclaimer: Vapi/Retell/LiveKit prices are published ranges as of July 2026, subject to change by plan / region / volume. Confirm in your provider's dashboard before quoting a customer.

What "GPT-Live full-duplex free" means for your unit economics

The headline says "zero cost" — but reality has three layers: Free with caps, included paid plan, and real production on the gpt-realtime-2.x API family. Here is the math for the most common case: a Spanish-language customer service SaaS agent for Latam.

Base case: SaaS with 10,000 calls/month of Spanish customer service, 2.5 min per call, 60% user / 40% agent speaking mix. That is 25,000 minutes/month of audio total (15,000 user minutes + 10,000 agent minutes).

Layer 1 — Demo and prototyping on GPT-Live Free

The Free plan lets you use GPT-Live-1 mini at no cost. The catch: the minute caps match ChatGPT Free and OpenAI changes them without much notice (Eesel). For a founder who needs 50–200 minutes/month for demos, A/B tests and customer validation, Free is enough and the marginal cost is zero. You do not pay Twilio because calls run through the ChatGPT app (you cannot use your number, but for a demo that does not matter).

Cost per layer: $0. The only thing you pay is your dev's time.

Layer 2 — Production on Plus/Pro plan

Plus ($20/month) carries included minutes but stays tied to the ChatGPT app — you cannot plug in your Twilio, your CRM or your analytics. For B2B SaaS in Latam, this layer is only useful if the use case is "user talks to a ChatGPT expert in my product" (a real pattern for onboarding).

Cost per layer: $20/month per human agent that logs in, no scalability to 100+ concurrent calls.

Layer 3 — Real production on the gpt-realtime-2.x API

The API brings the same full-duplex architecture to production. Pricing from the OpenAI pricing page captured July 13:

Math for 10,000 minutes/month in Spanish (60% user / 40% agent):

Audio to tokens conversion (Fora Soft): 1 minute of user audio = 600 tokens, 1 minute of agent audio = 1,200 tokens. For 25,000 total minutes (10K calls × 2.5 min): 9M input tokens, 12M output tokens, 8M cached tokens (2K system prompt repeated across 4K calls).

gpt-realtime-2.1-mini with prompt caching (most realistic Latam scenario):

• Fresh input: 9M × $10/1M = $90
• Cached input: 8M × $0.30/1M = $2.40
• Output: 12M × $20/1M = $240
• Total: $332/month → $0.0133/min all-in ($332 / 25,000 min)

gpt-realtime-2.1 flagship with prompt caching:

• Fresh input: 9M × $32/1M = $288
• Cached input: 8M × $0.40/1M = $3.20
• Output: 12M × $64/1M = $768
• Total: $1,059/month → $0.0424/min all-in ($1,059 / 25,000 min)

Compared with Vapi all-in ($0.13–$0.32/min): gpt-realtime-2.1-mini with cache is 90% cheaper than Vapi at the low end ($0.0133 vs $0.13) and 96% cheaper at the high end ($0.0133 vs $0.32). The flagship gpt-realtime-2.1 lands 67% cheaper than Vapi's low tier ($0.0424 vs $0.13) and 87% cheaper than the high tier ($0.0424 vs $0.32), with the added advantage of native 300 ms latency vs Vapi's 500–800 ms.

At 100,000 minutes/month (~700 calls/day, a mid-sized Latam SaaS volume) — 4x linear scale:

• gpt-realtime-2.1-mini with cache: $1,328/month
• gpt-realtime-2.1 flagship with cache: $4,236/month
• Vapi all-in equivalent: $13,000–$32,000/month

The gap of $12K–$31K/month between GPT-Realtime-mini and Vapi at scale pays for an extra engineer-month. Or, looked at the other way, it makes a $29/month plan viable when its previous inference COGS was 40%.

Calculation disclaimer: the tokens assume a 2,000-token system prompt repeated across calls. If your prompt is shorter, cache hit naturally drops and cost rises 10–20%. Run your own math with your real mix before committing.

What to audit in your SaaS this week

Operational checklist to run Monday to Friday. The idea: you have evidence for the next planning with the CFO, not intuition.

1. Identify your current voice minutes (or your intent to have them). If you already have a voice agent running, you need a dashboard showing minutes consumed by (model, day, channel, language) over the last 30 days. If you do not have voice yet, define the use case and estimate ticket × duration.

2. Calculate the cost per logical use, not per minute. "Support call resolved by AI agent" costs X. "Call transferred to human" costs X + Twilio transfer fee. "Voice-guided onboarding" costs W. That gives you unit COGS. If a Pro user pays $29/month and you generate $11 of voice for them (40 min × $0.28 all-in on Vapi), you have 38% COGS. If you migrate to gpt-realtime-2.1-mini with cache, COGS drops to ~$0.53 (1.8%, 40 min × $0.0133) — the same user is profitable again.

3. Measure your cross-border latency today. GPT-Live and gpt-realtime-2.x run on US infrastructure. From São Paulo, Buenos Aires or Bogotá expect a base RTT of 120–180 ms to the endpoint + 150–250 ms of model processing. Total: 300–430 ms, within the good range. The killer is when you add an STT (Deepgram) US layer + LLM (OpenAI) US + TTS (ElevenLabs) EU — the pipeline adds up to 500–800 ms. Native full-duplex makes that problem disappear.

4. Run a 100-call pilot on Free before paying anything. Activate GPT-Live Free, define a test script, call 5 friends with a dev's phone and ask them to interrupt, ask ambiguous questions and speak with an accent (rioplatense, Andean, costeño). Document: stutters on pauses, words not understood, response time, "person" vs "bot" feel. With 100 calls you have evidence to decide. Investment: 2–3 afternoons.

5. Negotiate commit with your provider past 50K minutes/month. Voice discounts are 20–35% off public list. On top of caching and the mini tier, you can end up paying ~$0.009/min instead of $0.0133/min — about 30% less than the cached list rate.

When NOT to use GPT-Live

GPT-Live is competitive but not a universal replacement. Three cases where Vapi, Retell, LiveKit or a classic IVR are still the right call:

1. Compliance and sensitive data. GPT-Live consumer (Free, Plus, Pro) is not available in Business, Enterprise or Edu workspaces (OpenAI Help Center via Digital Applied). If your SaaS sells to hospitals, financial institutions or government, Free is out by default. Vapi has a HIPAA add-on ($1K/month, per Retell) and Retell/LiveKit have native SOC2. GPT-Live on the API will have an enterprise path, but it has not been announced with a timeline.

2. You need an open and auditable tool stack. GPT-Live is a closed model: you pass in audio, get audio back, you cannot inspect the intermediate transcription or the tool call it chose. If your use case requires explainability (regulatory audit, production error debugging, conversation compliance), an open STT → LLM → TTS pipeline (Vapi, Retell, LiveKit) gives you access to every step. This matters in collections, health and legal.

3. Low volume with regional numbers. If you do 2,000 minutes/month and need numbers in Mexico, Argentina, Brazil and Colombia, the cost of porting numbers in Twilio + GPT-Realtime API is higher than paying Vapi all-in with numbers included. Vapi and Retell solve multi-regional; GPT-Live API has not announced regional presence yet.

When NOT to migrate to an expensive wrapper "just in case": if your case is customer service with a short script + FAQ, and you are above 50K+ minutes/month, the savings of migrating from Vapi to gpt-realtime-2.x with cache (90–96% by tier, mini against Vapi low/high) are material. Do not overpay for bundling if you can operate the stack yourself.

Three operational practices that apply wherever you are

Regardless of the stack you pick, three optimizations cut the bill 40–70% combined:

1. Turn on Voice Activity Detection (VAD) and prompt caching before switching providers. VAD prevents the model from spending tokens on user silence. Prompt caching drops the cost of the repeated system prompt to $0.30/MTok on the mini tier — 30x cheaper than fresh input. Without touching the model, these two changes drop the typical bill 50–60% (Aireiter).

2. Measure WER on your own accent set before promising an SLA to the customer. Rioplatense/neutral Spanish works well in OpenAI demos. Andean Spanish, Caribbean Spanish and Brazilian Portuguese with their phoneme mix are a different story. Record 30 audios from your real customers with varied accents, run them through the GPT-Live or Deepgram transcription, and measure WER (Word Error Rate). If your WER exceeds 12% in a segment, do not promise an SLA in that vertical — fall back to human when confidence drops below a threshold.

3. Design the human fallback from day one. Voice-first is not "replace humans" — it is "handle 80% of low-context calls and route the critical 20% to a human with context already loaded". A voice agent without a fallback degrades customer experience when it gets things wrong. Design the handoff from day one: timeout → human, transcription error → human, cancellation intent → human. The line between "voice-first works" and "voice-first destroys your brand" runs through that fallback.

Conclusion

The headline — "full-duplex voice on Free" — is real, but the naive read ("replace my call center with ChatGPT") is dangerous. What GPT-Live really changes: the cost of prototyping a voice agent drops to zero, and production unit economics improve 90–96% versus an all-in wrapper if you operate the Realtime API yourself. What it does NOT change: you need a wrapper (Vapi, Retell, LiveKit) for compliance, regional numbers and auditable tool stack, and you need a human fallback.

Three operational recommendations to close out July:

1. Run a 100-call pilot on GPT-Live Free this week. Not on production, not with real customers, not with your Twilio number. With 5 friends, a list of awkward questions and a stopwatch. If it passes the "feels like a person" test, you move to the next stage. If not, the problem was not the price.
2. Calculate your unit COGS by channel (voice / text / email / chat) before deciding the stack. If your voice COGS exceeds 30% of user revenue, you have a pricing or duration problem. No model fixes that for you if the agent stays 4 minutes per call.
3. If you sell to enterprise with compliance, wait for GPT-Live's enterprise API. That is quarters, not weeks. In the meantime, Vapi HIPAA or Retell SOC2 cover you. Do not put a $50K/year deal at risk to save $3K/month on provider.

If your startup is evaluating building a voice agent in Spanish/Portuguese or migrating from Vapi/Retell to the OpenAI API, book a free 30-minute call — in 20 minutes we can usually calculate your monthly savings with your real minute mix and use case.

---

Read also:

• Case study: how a Latam SaaS automated support and saved 380 hours/month — example of voice + chat + email in production, with numbers.
• Claude Sonnet 5 at $2/$10: review your COGS this week — if your voice agent delegates to a reasoning LLM, that cost line also changed in July.
• How to choose an AI consultant: 7 questions before signing — framework if you are outsourcing the voice agent.
• Tutorial: SaaS support agent with LangGraph — the "brain" behind the voice agent, step by step.
• Back to the blog — all articles.

Frequently asked questions

What is GPT-Live and how is it different from ChatGPT Voice?

GPT-Live is OpenAI's new family of full-duplex voice models. Unlike ChatGPT Voice (turn-based), it listens and speaks at the same time: it lets you interrupt, does not wait for the end of your turn, and keeps natural prosody and pauses. It is rolling out to Free, Go, Plus and Pro.

How much does GPT-Live really cost in production from Latam?

On the Free plan the monetary cost is zero but there are minute caps. Plus/Pro carry included minutes. On the API, pricing is not yet published for all regions and you still have to add cross-border latency (US servers), which in voice is a killer. For a Spanish sales agent, an 800ms round-trip is the line between usable and frustrating.

Can GPT-Live replace my IVR or Retell/Vapi today?

Depends on the case. For a customer service agent with a short script and FAQ, yes — you can prototype in an afternoon. For an agent that negotiates, handles objections or passes sensitive data, you still need a wrapper (Vapi, Retell, LiveKit) that gives you control of barge-in, timeouts, transcription and compliance. GPT-Live is the model, not the stack.

What about Spanish/Portuguese quality and Latam accents?

OpenAI trained on multilingual data and rioplatense/neutral Spanish works well in demos. Andean and Caribbean accents are still the achilles heel — measure WER (word error rate) on your own set before promising an SLA to the customer. Have a human fallback when confidence drops below a threshold.