Microsoft's MAI Realtime 1 Beats GPT Voice at One Cent per Minute
Microsoft's first speech-to-speech model undercuts GPT Realtime 2.1 on price and claims higher accuracy, with OpenAI Realtime API compatibility for easy migration.
Takeaways
- Microsoft launched MAI Realtime 1, its first in-house speech-to-speech model, at Ignite 2026.
- Pricing starts at $0.01 per audio minute, plus $1/$5 per 1M input/output reasoning tokens.
- Compatible with the OpenAI Realtime API, enabling drop-in migration of existing voice apps.
- Reports 98.9% accuracy on Big Bench Audio and claims to beat GPT Realtime 2.1.
- Tiered architecture: fast core for conversation, heavier reasoning engine for complex tasks.
- Available in MAI Labs and limited-access Foundry preview; Agoda and Reservations.ai already testing.
Microsoft has introduced MAI Realtime 1, its first native speech-to-speech model. The preview model starts at $0.01 per audio minute, adds token charges for inference, and supports the OpenAI Realtime API. Microsoft claims it outperforms GPT Realtime 2.1 on quality, speed, and price, giving developers an OpenAI-compatible option for testing voice agents.
Realtime 1 expands Microsoft’s in-house MAI model family into live conversation. Another model in that portfolio, MAI-Transcribe-2, powers 10 million monthly patient encounters in Dragon Copilot, according to Microsoft. Realtime 1 adds full-duplex audio, which allows a system to listen and speak within the same session while handling interruptions.
Speech stays in one loop
Traditional voice agents chain together speech recognition, text generation, and speech synthesis. Each handoff adds latency and can discard timing, tone, and emphasis from the original audio. Realtime 1 processes audio input and output within a unified conversational loop designed to preserve those signals.
Microsoft describes a tiered architecture in which a fast core handles routine conversation and routes harder requests to a stronger reasoning engine. This design aims to deliver quick acknowledgments while reserving more computation for complex questions. The release leaves the routing thresholds and the identity of the reasoning engine undisclosed.
Interruptions, noise, and tools
- Turn management: Detects when a speaker has finished, handles interruptions, and produces short acknowledgments.
- Adaptive reasoning: Uses a fast path for routine exchanges and a stronger reasoning path for complex tasks.
- Noisy-audio handling: Processes requests containing background speech and ambient sound.
- Tool calling: Emits calls that a host application can route to enterprise APIs and services.
- Prompted personas: Adjusts tone, style, and conversational behavior through instructions.
One cent, plus token charges
Microsoft’s announced preview pricing separates audio processing from inference tokens:
| Usage | Price |
|---|---|
| Speech-to-speech audio | $0.01 per audio minute |
| Input inference | $1 per million tokens |
| Output inference | $5 per million tokens |
At the published rate, 10 minutes of billable audio costs 10 cents before inference, tools, storage, networking, and application infrastructure. The release leaves unclear whether simultaneous input and output streams are metered separately. Total cost will also depend on conversation length, prompt size, tool results, and how often the model invokes deeper reasoning.
Strong claims, sparse evidence
Microsoft says Realtime 1 beats GPT Realtime 2.1 across Big Bench Audio, τ-Voice, and internal Dynamics 365 Contact Center evaluations. Its model page reports 98.9% accuracy on Big Bench Audio and places the system on a price-performance Pareto frontier, meaning Microsoft’s comparison found no tested model with a better score at the same or lower cost.
| Evaluation | What it measures | Published evidence |
|---|---|---|
| Big Bench Audio | Reasoning over spoken inputs | 98.9% accuracy |
| τ-Voice | Agentic task completion by voice | Outperformance claim without a score |
| Dynamics 365 evaluation | Contact-center performance | Internal result with unpublished methodology |
The public evidence remains limited to one accuracy result, an internal latency figure, and Microsoft’s comparative claims. Independent replication has yet to appear. Microsoft has also withheld a model card, supported-language list, voice catalog, context window, regional availability, and rate limits, leaving several production constraints unresolved.
Compatibility lowers migration work
Support for the OpenAI Realtime API may allow existing applications to retain their event-driven session architecture and much of their client code. An existing client should still expect to configure a new endpoint, credentials, model identifier, and provider-specific options.
Protocol compatibility can coexist with different behavior in turn detection, interruption handling, audio codecs, tool-call schemas, error events, and session limits. An A/B test should compare those details alongside response quality, task completion, latency, and cost.
Where Microsoft is testing it
- Agoda and Reservations.ai are running trials for booking assistants.
- Dynamics 365 Contact Center plans to use the model in customer-service applications.
- Dragon Copilot plans to use it for simulated patient conversations that help nurses practice clinical communication.
Preview access comes in two forms
| Channel | Access | Billing |
|---|---|---|
| MAI Labs | Interactive trial available to anyone | No production API terms stated |
| Microsoft Foundry | Registration-based limited preview | No charges until public preview begins |
Because Realtime 1 remains in preview, Microsoft has yet to announce a general-availability date, production service-level agreement, regions, or final rate limits. The published prices and technical terms may also change before release.
A broader first-party stack
At Ignite on 17 November 2026, Microsoft announced Realtime 1 as part of a broader batch of first-party models spanning several workloads:
- Software development: MAI Code 2 and MAI Code 2 Flash
- Cybersecurity: MAI Cyber 2 Flash
- Image generation and editing: MAI Image 2.7 and its Flash variant
- Real-time voice: MAI Realtime 1
- Clinical transcription: Dragon Copilot Transcribe, built on MAI-Transcribe-2
- Decision modeling: Microsoft-Decision-1.1
Adding real-time speech gives Microsoft a first-party option in a category where it also integrates partner models, including OpenAI’s. The expanded portfolio lets Microsoft offer its own models across coding, security, images, transcription, decision systems, and live voice.
A production test checklist
A production trial should use representative calls and measure the complete application path, including network transport, tool execution, and response playback:
- Task success: Completion rates across supported languages, accents, speaking speeds, and noisy environments.
- Latency: Median and tail latency for first audio, complete responses, and tool-assisted turns.
- Turn-taking: False interruptions, missed barge-ins, long pauses, and overlapping speech.
- Tool reliability: Argument accuracy, timeout handling, retries, and recovery from failed calls.
- Total cost: Audio minutes, input and output tokens, tools, storage, networking, and observability.
- Operational fit: Data retention, regional processing, compliance terms, rate limits, and preview support.
The low base rate and API compatibility reduce the engineering and usage cost of a parallel evaluation. Production readiness will depend on workload-specific results and on the limits, languages, service terms, and safety documentation Microsoft has yet to publish.