On May 7, 2026, OpenAI dropped three new models in its Realtime API. One of them, GPT-Realtime-Translate, streams voice translation from 70 input languages into 13 output languages — including English and French — for $0.034 per minute.
That's $2.04 per hour of translated conversation, running flat out.
In Montreal, the first reflex was predictable: "OK, let's bolt this onto our AI voice agent and serve all of Canada without hiring." Except the answer isn't that simple. Bolting GPT-Realtime-Translate onto an agent that already speaks French and English natively means adding a layer — meaning extra cost, latency, and a new point of failure — to solve a problem you may not actually have.
This article gives you the 6-signal decision framework. If three or more light up for your SMB, the bolt-on makes sense. If fewer, keep your current bilingual architecture.
What OpenAI Actually Shipped on May 7, 2026
First, the facts — because half the LinkedIn posts this week conflate the three new models. To dig into the technical detail, OpenAI published the official announcement on May 7.
GPT-Realtime-Translate is:
- 70 input languages, 13 output (English, French, Spanish, Portuguese, Japanese, Russian, Chinese, German, Korean, Hindi, Indonesian, Vietnamese, Italian)
- Pricing: $0.034 per minute of audio (vs. $32 per million audio input tokens for GPT-Realtime-2)
- A dedicated model — trained on thousands of hours of professional interpreter audio
- "Pace-matched" behavior: it waits for enough context before producing the translated voice
This is not a conversational agent. It's an interpreter. It doesn't reason, it doesn't book appointments, it doesn't consult your RAG. Its single job is to translate voice in real time, keeping the speaker's pace. See the official implementation guide in the OpenAI cookbook.
Simple parallel: GPT-Realtime-2 is your receptionist who decides. GPT-Realtime-Translate is the interpreter who translates what's being said.
The Architecture Mistake to Avoid
The tempting picture when you read the announcement is a sandwich: customer speaks Spanish → translate to English → GPT-Realtime-2 processes → English reply → translate to Spanish → customer hears.
In theory, it works. In practice, you've just stacked two models, doubled latency (each hop adds 200-400 ms), pay twice and created two failure points. For a 5-minute call with a Spanish-speaking customer: $0.17 for translation plus the normal GPT-Realtime-2 cost.
For a Quebec AI voice agent that mostly serves French and English, native bilingual architectures already exist without needing a translation layer. The conversational model handles both languages without a relay. Faster, simpler, cheaper.
So the real question isn't "Will GPT-Realtime-Translate improve my agent?". It's "Does my agent see enough calls in languages my base model can't handle to justify adding a layer?".
That's where the six signals come in.
The 6 Signals That Justify the Bolt-On
Signal 1 — Your inbound calls fall outside the FR/EN pair more than 5% of the time
A natively bilingual ElevenLabs or GPT-Realtime-2 agent handles Quebec French and North American English without any help. The model switches between them on the fly based on what the caller speaks. No translation layer needed. But if you're a Montreal retailer seeing 8% of calls in Spanish (US tourists), 4% in Mandarin (Chinese customers) and 2% in Portuguese, those 14% go to voicemail — or worse, hang up. At 14% of 800 calls/month with an average ticket of $120, that's monthly losses around $4,000. The bolt-on becomes profitable in the second month.
How do you know your actual distribution? Pull 90 days of call records and have someone listen-tag the first 20 seconds of each. Tedious but cheap. Or have your existing voice agent log the language detected by its STT and run a query at the end of the quarter. Either way, you need data — not a gut feeling.
Signal 2 — You have a strong tourist season
Inns, hotels and restaurants in Old Quebec and Mont-Tremblant see their language ratio explode between May and October. If that's you, the read for you is on multilingual deployment in the Quebec hospitality sector. Plugging in translation seasonally (May-October) then disabling it off-season cuts the annual bill in half.
Signal 3 — You handle regular cross-border business
Immigration law firms, customs agencies, exporters to Latin America or Asia. Volume isn't the point here, stakes are: one missed call worth $20,000 in contract value covers many hours of translation at $2.04/hour. The conversation that opens a partnership doesn't always come in English — and asking a prospective client to "call back later when our bilingual associate is in" is the kind of friction that pushes them straight to your competitor.
Signal 4 — Your market contains an identifiable diaspora
Medical clinics in Côte-des-Neiges (strong Filipino community), pharmacies in Brossard (Vietnamese clientele), garages in Saint-Léonard (Italian). If more than 10% of your customers prefer to speak in their original language, the customer experience justifies the cost. Those 10% often include your most loyal customers — they stay where they're served in their language.
Signal 5 — You want to document Bill 96 compliance without bloating the agent
Section 5 of the Charter of the French Language: customer communications must be in French unless the customer explicitly asks otherwise. GPT-Realtime-Whisper (the third model shipped May 7, at $0.017/min) creates verifiable transcriptions of calls. Plugged in alongside your existing agent, it provides the written trail the Office québécois de la langue française could request, without changing a word of the customer flow.
Signal 6 — Your marginal cost per call supports a 2-3% increase
Cold math: if your calls currently cost $0.08/minute (LLM + TTS + STT), adding $0.034 is a 42% jump. If your calls cost $0.80/minute (premium agents like GPT-Realtime-2 at high reasoning levels), it's a 4% jump. What you were already paying entirely determines whether the translation layer is negligible or painful. To understand how to tune the reasoning levels and therefore the base cost, the article on GPT-Realtime-2's 5 reasoning levels walks through it.
When NOT to Bolt It On (the part nobody says)
If you check fewer than three signals, you're wasting budget. Three concrete cases where we advised against the bolt-on in May:
- Sherbrooke dental practice: 99% French calls, 1% English. The native model already handles English — adding Translate is paying for nothing.
- Longueuil plumber, 50 calls/month: volume too low to amortize integration cost. The math comes out under $12 per translated call — not justifiable.
- Family restaurant on the Plateau: 95% local clientele. The 5% English-speakers are already served by the model's native bilingualism. Translation adds perceptible latency (200-400 ms) that degrades experience for 100% of calls to serve nobody more.
The unwritten rule: if the translation layer improves fewer than 1 in 20 customers, it degrades the experience of the other 19.
How TECHMA Deploys This for You
All GPT-Realtime-Translate bolt-ons to existing AI voice agents go through the TECHMA team — you don't touch code or manage API keys. The standard process takes 5-7 business days: analysis of your language distribution over the past 90 days, break-even calculation, configuration of conditional activation (e.g., only fire Translate if language detection identifies a non-FR/EN language), sandbox testing, then production cutover with monitoring.
The TECHMA integration cost is included in our plans — no surprise overcharges when the API evolves.
FAQ
Do I need to replace my current agent?
No. GPT-Realtime-Translate plugs in as a conditional layer. Your ElevenLabs, Vapi, or Retell agent keeps running. Translation only fires when language detection triggers the need.
Is the latency added by the translation layer perceptible?
Yes. Count 200-400 ms additional per turn of speech. On a call in the target language, the experience remains fluid. On an English-French call routed through translation instead of native bilingualism, you lose responsiveness — hence the conditional activation.
What's the typical payback period?
For SMBs that check at least 3 signals, payback is 2-4 months. For those checking all 6, it's immediate — the first saved call often pays for the year.
And privacy — do calls go through OpenAI?
Yes, as with GPT-Realtime-2. OpenAI retention policies apply. For medical clinics covered by Quebec Law 25, TECHMA sets up data processing agreements and enables no-retention options. The details on OpenAI's official pricing are here.
My agent supports 8 languages natively — do I still need Translate?
No, unless you drop to a lower reasoning level for cost reasons and want translation as an optional premium layer for select VIP calls.
How long before Translate is superseded?
Given OpenAI's pace (three models shipped in May 2026), expect a 6-9 month cycle before the next iteration. Your integration is automatically updated by the TECHMA team.
Bottom Line
Three signals or more: deploy. Fewer than three: save and focus your budget on native agent quality. The worst scenario is adding a layer because it's trendy and degrading the experience of 95% of your calls to serve 5%.
For a quick analysis of your actual language distribution, request a free read with Agent IA Vocal — we look at your last 90 days of calls and decide in 20 minutes.