Voice agent costs in 2026: MAI-Voice-2-Flash at $15 per 1M characters vs OpenAI Realtime at $64

Summary. Microsoft priced MAI-Voice-2-Flash at $15 per 1M characters on 23 July 2026, 32% below MAI-Voice-2 and twice as fast. OpenAI's gpt-realtime-2.1, released in July 2026, bills $32 per 1M audio input tokens and $64 per 1M audio output tokens. Those two price lists are quoted in different units, which is why most build-versus-buy spreadsheets get voice economics wrong. Converted to a common basis using Microsoft's own published token rates for Azure Voice Live (about 10 audio tokens per second inbound, 20 outbound) and ElevenLabs' own equivalence of 1,000 characters to roughly one minute of speech, a six-minute support call costs about $0.288 on gpt-realtime-2.1 and about $0.063 on a Microsoft Foundry cascade of MAI-Transcribe-1.5 plus MAI-Voice-2-Flash. At 100,000 calls a month that is $28,800 against $6,300, a gap of roughly $22,500 per month, or about ₹21.4 lakh at the 3 August 2026 rate of ₹95.29 to the dollar.

The gap is real, but it is not free money. The cheaper number buys you an orchestration problem, three vendor SLAs instead of one, and a latency budget you now own. This article does the arithmetic, shows the assumptions, and marks the point where the expensive option is still the right call.

The two architectures you are actually choosing between

Every production voice agent in 2026 is one of two shapes.

The cascade runs speech-to-text, then a text LLM, then text-to-speech, as three separate calls per turn. You pay for transcription by the audio minute or hour, for the language model by the text token, and for synthesis by the character. Each stage is swappable. Each stage adds latency.

The speech-to-speech model takes audio in and emits audio out inside one model, with no text round trip in the middle. OpenAI's Realtime API is the reference implementation. Azure Voice Live exposes the same shape on Microsoft's side, and Microsoft says MAI-Voice-2-Flash is now integrated into Voice Live so developers can build speech-to-speech agents there too.

Speech-to-speech wins on prosody and interruption handling because the model hears tone, pauses and overlap directly. The cascade throws that away at the transcript boundary. What the cascade wins is price control, model choice per stage, and the ability to log a clean text transcript for evaluation and compliance, which matters more than teams expect at audit time.

Luke Miller, CEO of the UK voice infrastructure startup SLNG, put the cost dynamic bluntly to Tech.eu in May 2026: "The market has been shaped by voice labs whose business model depends on maximising compute at every step of every call. Every syllable through the most expensive TTS, every pause analysed by a full LLM call, every transcription through the highest-cost engine. We want to reprice the entire voice agent market." SLNG says teams routing through its execution layer see model costs drop by over 50% and per-turn latency cut by more than half.

What each vendor actually charges, in its own units

These are published list rates as of 4 August 2026, each in the unit the vendor bills in.

Component	Vendor and model	Published rate
Text to speech	Microsoft MAI-Voice-2-Flash (public preview)	$15 per 1M characters
Text to speech	Microsoft MAI-Voice-2	32% above Flash, so about $22 per 1M characters
Text to speech	ElevenLabs Flash / Turbo	$0.05 per 1,000 characters
Text to speech	ElevenLabs Multilingual v2 / v3	$0.10 per 1,000 characters
Speech to text	Microsoft MAI-Transcribe-1.5	$6 per 1,000 minutes of audio
Speech to text	ElevenLabs Scribe v2	$0.22 per hour
Speech to text	ElevenLabs Scribe v2 Realtime	$0.39 per hour
Speech to speech	OpenAI gpt-realtime-2.1	$32 in / $64 out per 1M audio tokens
Speech to speech	OpenAI gpt-realtime-2.1-mini	$10 in / $20 out per 1M audio tokens
Bundled pipeline	ElevenLabs Speech Engine	$0.08 per minute

Five different billing units across ten rows. That is the whole problem. A procurement team comparing "$15" to "$64" reads a 4x gap; the real per-minute gap is different again, and in one direction it is larger.

Converting everything to cost per minute

Two conversion factors do all the work, and both come from vendor documentation rather than estimation.

Audio tokens per second. Microsoft's Voice Live API documentation publishes a token estimation table: for Azure OpenAI models, input audio is about 10 tokens per second and output audio about 20 tokens per second (Phi models bill input at about 12.5). That gives 600 input tokens and 1,200 output tokens per minute of speech.

Characters per minute of speech. ElevenLabs' own API pricing page prices Flash at $0.05 per 1,000 characters and annotates the same row as approximately $0.05 per minute, so the vendor itself treats 1,000 characters as about one minute of synthesised speech.

Apply both:

# Speech-to-speech, per minute of audio
caller audio in  = 600 tokens  x  $32 / 1,000,000  = $0.0192
agent audio out  = 1,200 tokens x $64 / 1,000,000  = $0.0768

# Cascade, per minute of audio
caller audio in  = 1 min x $6 / 1,000 min          = $0.0060
agent audio out  = 1,000 chars x $15 / 1,000,000   = $0.0150

Per minute of audio	gpt-realtime-2.1	gpt-realtime-2.1-mini	MAI cascade	ElevenLabs cascade
Caller speech (inbound)	$0.0192	$0.0060	$0.0060	$0.0065
Agent speech (outbound)	$0.0768	$0.0240	$0.0150	$0.0500
Ratio to MAI cascade, outbound	5.1x	1.6x	1.0x	3.3x
Text LLM billed separately	No	No	Yes	Yes
Single vendor SLA	Yes	Yes	Yes	Yes

Outbound audio is where the money goes, because the agent talks more than the caller in most support flows and because output tokens cost twice input tokens on the Realtime API. Synthesising a minute of agent speech costs $0.0768 on gpt-realtime-2.1 and $0.0150 on MAI-Voice-2-Flash. That is the 5x.

A 100,000-call month, priced four ways

Model a six-minute call with a 50/50 speaking split: three minutes of caller audio, three minutes of agent audio. This is deliberately conservative, since real support calls skew toward the agent.

Stack	Per call	100,000 calls / month	Monthly in ₹ (at ₹95.29/$)
gpt-realtime-2.1	$0.288	$28,800	₹27.4 lakh
ElevenLabs Speech Engine (bundled)	$0.480	$48,000	₹45.7 lakh
gpt-realtime-2.1-mini	$0.090	$9,000	₹8.6 lakh
ElevenLabs Scribe v2 Realtime + Flash TTS	$0.170	$17,000	₹16.2 lakh
MAI-Transcribe-1.5 + MAI-Voice-2-Flash	$0.063	$6,300	₹6.0 lakh

The spread between the cheapest and dearest row is $41,700 a month on identical call volume. Rupee figures use the 3 August 2026 USD/INR rate of ₹95.29 reported by Trading Economics.

Three caveats before anyone takes that table to a budget meeting.

The cascade rows exclude the text LLM. Microsoft's Voice Live example pricing for East US lists text input at $4.40 per 1M tokens, cached text input at $1.375 per 1M, and text output at $17.60 per 1M. A six-minute call generating, say, 2,000 text output tokens adds roughly $0.035 per call at that rate, which moves the MAI cascade from $0.063 to around $0.098. Still under a third of gpt-realtime-2.1, but the honest number is not $0.063.

The speech-to-speech rows exclude context re-billing. On a stateful realtime session the accumulated audio context is re-sent on each turn unless it is cached. OpenAI prices cached input at $0.40 per 1M tokens for gpt-realtime-2.1, a 20% reduction on the older gpt-4o-realtime-preview, which is exactly the lever that keeps long calls affordable. Treat the per-minute figures above as a floor, not a forecast, and instrument your own sessions before committing.

Nobody pays list at volume. All four vendors negotiate, and MAI-Voice-2-Flash and MAI-Image-2.5-Pro were both still in public preview as of 23 July 2026, so preview pricing can move.

What the accuracy and latency benchmarks say

Price only matters if quality holds. Independent evaluation from Artificial Analysis on 2 June 2026 placed MAI-Transcribe-1.5 third overall on its word error rate leaderboard at 2.4% AA-WER, behind Alibaba's Fun-Realtime-ASR-preview at 1.7% and ElevenLabs Scribe v2 at 2.2%. Its distinguishing number is speed: a processing factor of about 276x real time, more than double the second-fastest model in the accuracy top 10. It ranked second on VoxPopuli-Cleaned-AA at 1.6% WER, fourth on Earnings22-Cleaned-AA at 4.0%, and fifth on AA-AgentTalk at 2.0%.

On the synthesis side, Microsoft reports MAI-Voice-2 was preferred over MAI-Voice-1 in 72.1% of 2,500 side-by-side listening tests. In a separate evaluation across 11 languages with 2,222 responses, 45.5% of listeners preferred MAI-Voice-2 output, 44% preferred real human recordings, and 10.5% called it a tie. MAI-Voice-2-Flash is described as retaining that acoustic quality while running twice as fast.

For end-to-end latency, DestiLabs published a 2026 voice agent benchmark drawn from more than ten production projects reporting median end-to-end response latency of 680 ms at p50 and 1,180 ms at p95, with all-in cost of $0.07 to $0.21 per minute depending on architecture and call volume. Those numbers bracket the per-minute figures above once orchestration, telephony and LLM cost are included, which is a useful sanity check on any vendor quote that lands far outside the band.

Metric	MAI cascade	gpt-realtime-2.1	Source of figure
Transcription accuracy	2.4% AA-WER (#3)	Not separately published	Artificial Analysis, June 2026
Transcription speed factor	~276x real time	Not applicable	Artificial Analysis, June 2026
Synthesis preference	72.1% over MAI-Voice-1	Not separately published	Microsoft AI, June 2026
Language coverage, synthesis	15 languages plus locales	Not separately published	Microsoft AI, June 2026
Turn latency ownership	Yours to engineer	Vendor-managed	Architectural

The blank cells are the point. OpenAI does not publish component-level WER or synthesis preference scores for gpt-realtime-2.1, because there are no components. That is a genuine trade: you buy an integrated experience and give up the ability to attribute a regression to a stage.

Where MAI-Voice-2 earns its price outside the cost table

Three capabilities are hard to price and easy to underestimate.

Zero-shot voice prompting builds a custom brand voice from 5 to 60 seconds of reference audio with no retraining or fine-tuning, across every supported language. Microsoft enforces consent at the system level, so only authorised, licensed voices can be synthesised in production and unlicensed cloning is blocked. For a regulated buyer, that guardrail is a procurement answer, not a feature bullet.

Emotion control uses explicit tags such as sad, whispered and excited, which gives a deterministic handle on delivery instead of hoping the model infers tone from context.

Code-switching is supported for selected language pairs, notably Hindi-English and Spanish-English. Microsoft's published sample switches mid-sentence without losing speaker identity or prosodic naturalness. For Indian customer support that is not a nice-to-have.

Microsoft also has production references rather than demos. MAI-Voice-2-Flash powers Dynamics 365 Contact Center, used by T-Mobile and EasyJet, and Microsoft reports it cut GPU costs there by up to 89%. MAI-Transcribe-1.5 replaced the previous model inside Dragon Copilot, a clinical documentation product used by 170,000 medical providers that processed 28 million patient encounters in a quarter, with Microsoft reporting a 50% relative reduction in transcription and language-identification error rates across most languages in internal evaluations.

When speech-to-speech is still the right answer

Pick gpt-realtime-2.1 or an equivalent when the conversation is adversarial, emotional or heavily interrupted. Debt collection, medical triage, crisis lines and high-value sales calls all depend on hearing hesitation and talking over the caller correctly. A cascade that transcribes, thinks and then speaks will lose that texture no matter how good each stage is.

Pick it when your team is small. The cascade saves $22,500 a month at 100,000 calls, and it costs you an engineer's ongoing attention on barge-in handling, endpointing, jitter buffers and failover between three providers. Below roughly 20,000 calls a month the saving does not cover that headcount.

Pick the cascade when call volume is high and flows are scripted: order status, appointment booking, KYC verification, delivery rescheduling. Those are the flows where SLNG's argument bites hardest, because routing every scripted turn through a frontier model is paying for judgement you do not need.

The real cost is usually the orchestration, not the model.

India-specific considerations

Data residency is the first constraint, not the last. Under the Digital Personal Data Protection Act 2023, voice recordings of Indian customers are personal data, and call audio that includes account numbers, health details or identity documents raises the stakes on where inference happens. Both Microsoft Foundry and the major speech vendors offer regional deployment, but the default endpoint is rarely the Indian one, and a voice agent assembled from three vendors has three residency questions rather than one.

SLNG built sovereign infrastructure across 11 regions including India, Singapore, Indonesia and the UAE precisely because of this, and Miller's framing is worth repeating: "A voice agent handling a mortgage application in Australia can't have its audio processed in Virginia. A patient triage system in Switzerland can't send recordings to a US-hosted model. These aren't edge cases, this is where the enterprise demand is." Ixigo, India's second-largest online travel agent, moved a significant portion of its customer support onto that layer.

Language mix is the second constraint. Indian support calls are rarely monolingual. MAI-Voice-2's Hindi-English code-switching addresses the synthesis half, and MAI-Transcribe-1.5's 43-language coverage addresses part of the transcription half, but neither covers the full set of Indian languages a national D2C brand needs. Teams building for that spread should read our note on multilingual voice agent options for India before assuming a single vendor closes the gap.

Cost sensitivity is the third. At ₹95.29 to the dollar on 3 August 2026, the $22,500 monthly gap between gpt-realtime-2.1 and the MAI cascade is about ₹21.4 lakh, which is real money against an Indian support operation's budget. It is also roughly what two senior voice engineers cost, which is why the build-versus-buy line sits where it does.

How to benchmark this yourself in a week

Vendor arithmetic is a starting hypothesis. Run the measurement.

Record 100 real calls from your own queue, anonymised, and measure the actual caller-to-agent speaking ratio and mean call duration. Most teams assume 50/50 and discover 35/65 in the agent's favour, which shifts the comparison further toward whichever stack has the cheaper outbound rate.

Count characters, not words, on your agent's actual scripted responses. A 1,000-character-per-minute assumption is a vendor's convenience; your prompts, disclaimers and confirmations may run denser.

Instrument token counts on a live realtime session for a full call rather than trusting the seconds-to-tokens conversion. The Voice Live guidance gives approximate rates, and audio encoding, silence handling and barge-in all move the real number.

Measure p95, not p50, for latency, and measure it from the caller's last syllable to the agent's first, not from your API call. The DestiLabs band of 680 ms p50 to 1,180 ms p95 is a reasonable pass mark.

Price the failure mode. A cascade with three vendors has three availability numbers that multiply. Decide in advance which stage degrades gracefully and which one drops the call.

For a deeper walk through the speech-to-speech side specifically, see our gpt-realtime-2.1 voice agent build guide, and for the full-duplex behaviour that separates a good agent from a demo, our note on full-duplex voice agents on the Realtime API. Both sit inside our broader guide to enterprise AI agents in production. If the decision is really managed-versus-build rather than model-versus-model, Amazon Connect's agentic voice option is the third path worth pricing.

FAQ

How much cheaper is MAI-Voice-2-Flash than the OpenAI Realtime API?

Per minute of synthesised agent speech, MAI-Voice-2-Flash costs about $0.015 against $0.0768 for gpt-realtime-2.1 audio output, roughly a five-fold difference. On a six-minute call with an even speaking split, a Microsoft Foundry cascade lands near $0.063 before the text model, against $0.288 for gpt-realtime-2.1.

What does MAI-Voice-2-Flash actually cost?

Microsoft priced MAI-Voice-2-Flash at $15 per 1M characters when it entered public preview on 23 July 2026. Microsoft describes it as twice as fast as MAI-Voice-2 and 32% cheaper, which places MAI-Voice-2 near $22 per 1M characters. Preview pricing can change before general availability.

How many audio tokens is one minute of speech on the Realtime API?

Microsoft's guidance for Azure Voice Live puts audio input near 10 tokens per second and audio output near 20 tokens per second, giving 600 inbound and 1,200 outbound tokens per minute. Treat that as an approximation and instrument a real session, because encoding and silence handling move the actual count.

Is a cascade always cheaper than speech-to-speech?

No. Below roughly 20,000 calls a month the saving does not cover the engineering time a three-vendor pipeline demands for barge-in handling, endpointing and failover. The cascade also excludes the text model, which on Microsoft's East US Voice Live rates adds around $0.035 per six-minute call.

How accurate is MAI-Transcribe-1.5 compared with the alternatives?

Artificial Analysis ranked it third overall at 2.4% AA-WER in June 2026, behind Alibaba's Fun-Realtime-ASR-preview at 1.7% and ElevenLabs Scribe v2 at 2.2%. Its advantage is speed rather than raw accuracy: a processing factor near 276x real time, more than double the next fastest model in the accuracy top ten.

Does MAI-Voice-2 handle Hindi and English in the same sentence?

Yes. Microsoft supports code-switching for selected language pairs including Hindi-English and Spanish-English, and reports that the model switches mid-sentence without losing speaker identity or prosodic naturalness. MAI-Voice-2 covers 15 languages plus additional locales, which is narrower than the language spread many Indian support operations need.

What latency should a production voice agent hit?

DestiLabs reported median end-to-end response latency of 680 ms at p50 and 1,180 ms at p95 across more than ten production deployments in 2026, with all-in cost of $0.07 to $0.21 per minute. Measure from the caller's last syllable to the agent's first, not from your own API call.

Can I clone a brand voice with MAI-Voice-2?

Microsoft supports zero-shot voice prompting from 5 to 60 seconds of reference audio across all supported languages, with no retraining required. Consent is enforced at the system level, so only authorised and licensed voices can be synthesised in production, and unlicensed cloning is blocked. Access requires an application to Microsoft.

How eCorpIT can help

eCorpIT builds and operates production voice agents for Indian and global teams, including the parts that do not appear on a pricing page: barge-in handling, endpointing, per-stage failover, transcript logging for evaluation, and data-residency architecture. We design applications aligned with Digital Personal Data Protection Act 2023 requirements, and we are CMMI Level 5, MSME Certified and ISO 27001:2022 certified. If you want the per-minute numbers above validated against your own call recordings before you sign a vendor contract, talk to our engineering team.

References

1. Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash — Microsoft AI, 23 July 2026.
2. Introducing MAI-Voice-2 — Microsoft AI, 2 June 2026.
3. MAI-Transcribe-1.5: new speech to text model leading the accuracy-speed Pareto frontier — Artificial Analysis, 2 June 2026.
4. Pricing for Azure Voice Live API — Microsoft Q&A, February 2026.
5. Voice Live API for real-time voice agents — Microsoft Learn, updated 2 June 2026 (token usage and cost estimation table).
6. ElevenAPI pricing — ElevenLabs, accessed 4 August 2026.
7. OpenAI releases gpt-realtime-2.1 and gpt-realtime-2.1-mini for low-latency voice agents — MarkTechPost, 6 July 2026.
8. OpenAI Realtime API pricing 2026: cost per minute math — Layer3Labs.
9. Introducing gpt-realtime and Realtime API updates for production voice agents — OpenAI.
10. Former Vercel executive says voice agents will be won on infrastructure, not models — Tech.eu, 18 May 2026.
11. 2026 AI voice agent benchmark: latency and cost per minute — DestiLabs.
12. Artificial Analysis speech to text leaderboard — Artificial Analysis.
13. Indian Rupee exchange rate — Trading Economics, 3 August 2026.

Last updated: 4 August 2026.