Microsoft previews MAI Realtime 1 with separate reasoning costs | Magica
Microsoft previews penny-a-minute MAI Realtime 1, with reasoning billed separately
Editorial Team
••📖5 min read
Microsoft’s first live speech-to-speech model offers OpenAI API compatibility and planned enterprise integrations. Its published evaluations show a low-cost option, with several rivals completing more agentic tasks.
MAI Realtime 1’s advertised base rate is $0.01 per audio minute; reasoning adds token charges.
Microsoft promises compatibility with OpenAI’s Realtime API and integrations into its own enterprise software.
Microsoft’s comparison tables show strong speech reasoning, but higher task completion from several more expensive rivals.
Microsoft AI, Microsoft’s in-house model development group, introduced MAI Realtime 1 on November 17, 2026, as its first live speech-to-speech model. The preview gives enterprises another voice-agent option from a supplier that also controls potential deployment channels in customer service and healthcare. Its low starting price leaves a practical question: what will a successfully completed conversation cost?
Live conversations build on Microsoft’s earlier voice work
MAI Realtime 1 combines listening, spoken responses and reasoning. Microsoft says it handles interruptions, brief acknowledgments and background noise, adjusts its persona through prompting, and connects to enterprise tools to carry out requests. A responsive system handles the live exchange and calls on a more powerful reasoning engine for harder tasks, according to its announcement.
Microsoft already had a speech-generation model. Its earlier MAI Voice 1 announcement described a system powering Copilot Daily and Podcasts. Realtime moves the company’s own lineup into live exchanges that combine listening, reasoning and action.
Those capabilities have competitive precedent. OpenAI, the developer of ChatGPT, introduced its Realtime API in October 2024 with speech-to-speech conversations, automatic interruption handling and function calling to trigger actions. An update on that announcement records general availability on August 28, 2025. Microsoft says compatibility with that API will help developers test MAI Realtime 1 and migrate applications; the work required for a particular application remains to be tested.
Microsoft is pitching Copilot Workspaces as a shared environment for people and AI agents. Team collaboration is already in early access, but personal Autopilot integration is still coming soon, with final agent pricing unannounced.
Editorial Team
The rivalry also runs through an existing financial and infrastructure relationship. In its April 2026 partnership update, Microsoft said it remained a major OpenAI shareholder and its primary cloud partner. Microsoft retained a license to OpenAI model and product intellectual property through 2032, now non-exclusive. Its own voice model therefore adds a competing supplier without, by itself, ending that partnership.
Company-reported τ-Voice task completion. Labels abbreviate Gemini 3.8 Live, Gemini 3.8 Live Extended Thinking (High), Phonic 1 (High) and GPT-Realtime-2.1 High. These benchmark scores do not measure customer-service resolution rates. Chart: Magica; data: Microsoft AI. Source: Microsoft AI model page.
A minute rate is not a completed-call price
Microsoft prices base speech-to-speech usage at $0.01 per minute. Reasoning adds $1 per million input tokens and $5 per million output tokens. Sixty billed audio minutes would therefore cost $0.60 before reasoning charges.
Its published model comparisons put MAI Realtime 1 at approximately $1.35 per hour of input audio for the evaluation workload. That normalized benchmark cost and the base minute rate measure different things. Neither establishes the total cost of a customer’s calls.
On speech reasoning, Microsoft reports 98.9% for its model on a fixed 40-question Big Bench Audio subset, compared with 95.8% for GPT-Realtime-2.1 High at $10.75 per input-audio hour. Microsoft attributes competitor scores and costs to Artificial Analysis, an AI benchmarking service, while supplying its own model’s result from internal evaluation. This is a specific benchmark comparison, not independent confirmation of Microsoft’s performance across customer workloads.
The same page’s τ-Voice comparison measures agentic task completion and includes only full-duplex models—systems that can listen and speak at the same time. It shows a more varied tradeoff:
Model configuration
Task completion
Cost per hour of input audio
Gemini 3.8 Live
30.1%
$0.84
MAI Realtime 1
57.5%
$1.35
Gemini 3.8 Live Extended Thinking (High)
68.6%
$3.50
Phonic 1 (High)
79.9%
$4.20
GPT-Realtime-2.1 High
45.7%
$10.75
These are figures published by Microsoft, with the same split between internally evaluated MAI results and competitor data attributed to Artificial Analysis. MAI beats the listed GPT-Realtime-2.1 configuration on both measures shown, but a cheaper alternative exists and several pricier configurations complete more tasks. The task-completion percentage is a benchmark result, not a measured customer-service resolution rate.
Distribution and infrastructure give Microsoft a route to customers
Microsoft says Agoda and Reservations.ai are testing MAI Realtime 1. It also plans integrations into Dynamics 365 Contact Center, its customer-service software, and Dragon Copilot, its clinical AI assistant. The Dragon use described is simulated patient conversations to help nurses practice clinical interactions.
Dragon Copilot already provides a deployment channel for another in-house model: Microsoft’s broader Ignite announcement says its separate Transcribe model powers 10 million doctor–patient encounters a month. That existing transcription volume is not adoption of Realtime 1.
Realtime is part of a seven-model release covering coding, security, images, voice, transcription and routing. Microsoft also says its second-generation Maia 200 AI accelerator is in production running its largest models. That gives context to its wider effort to supply models and computing infrastructure, but the announcement does not identify Realtime 1’s serving hardware or disclose its operating costs. Retail API prices alone cannot establish Microsoft’s margins.
The next test is performance on enterprise calls
Microsoft invites enterprise developers to register for a Foundry limited-access preview and says preview customers will not be billed until public preview begins. Its MAI Labs demonstration has a separate restriction: personal, noncommercial use only, excluding school and work activities. A consumer demonstration therefore does not establish enterprise deployment readiness.
The evidence buyers still need is completed tasks, response times and total charges on representative calls, including reasoning usage and any retries or human handoffs. The retained launch materials name testers but give no customer-specific completion rates or realized savings. Those results will determine whether Microsoft’s starting rate translates into an economical voice agent for a particular business.
Google is rolling out file references and sharing for Gemini skills as adult personal accounts lose the ability to use Gems. Automatic conversion preserves supported material, but file limits and missing tools can interrupt existing workflows.
Muddy Waters disclosed a short position in Nvidia while questioning a possible connection involving senior adviser Ajay K. Puri. Its own account leaves the key identity question unanswered, separating its new claims from an existing server-diversion prosecution.
Nvidia reported $115.8 billion in quarterly revenue in its first earnings release since co-founder Jensen Huang resigned. Strong demand gives interim CEO Colette Kress a financial cushion, while a campaign for Huang’s return collides with unresolved investigations and the board’s reported opposition.
OpenAI launched GPT-6.2 Astra and Sol for professional work, cutting Astra’s standard input and output token prices by 20%. The models offer cheaper paths through selected tasks, but research evaluations, reporting failures and unfinished independent safeguard testing complicate the case for broader autonomy.
Canberra-based Trellis Data is offering its 220-billion-parameter Agentaus 1 model to business and government. Its pitch goes beyond keeping data in Australia, but performance, deployment costs and a planned capital raising remain key tests.
Perplexity is adding direct collaboration between frontier models to Computer, letting a lead agent coordinate up to 16 supporting agents. Company-reported benchmarks suggest mixed teams can outperform repeated copies of one model, but leave customer costs and production reliability unresolved.
Marsh reported 105 client cyber incidents in the last two weeks of October, against 48 in the final two weeks of May. The broker links the increase partly to AI, while warning that insurance panel access may not guarantee an available response team.
Meta’s Muse personal AI agent has launched in Australia, with some local shopping and booking connections to follow. Its browser and payment tools broaden its reach, while the company’s technical disclosures qualify its privacy promises.
Microsoft and Salesforce are extending their sales software alliance with Copilotforce, combining CRM workflows with Microsoft 365 context under existing permissions. The Copilot plugin enters public preview, while Teams remains in pilot; earlier Microsoft tools and a Claude beta raise the bar for proving its value.
Microsoft expands MAI Image 2.7 from an editing preview into image generation and API access, with cheaper Flash pricing and a Canvas workspace. Benchmark claims depend on the workload, while the MAI Labs interface carries restrictions on commercial use.
OpenAI is widening applications for its legal AI configuration after improving research scores and expanding its search index. Access still requires approval, and benchmark results show substantial limits alongside the gains.
Microsoft has made Word, Excel and PowerPoint creation generally available inside Copilot, combining editable files with workplace permissions and collaboration. The broader shift toward autonomous work remains staggered across previews, while its claimed cost advantage rests on company-funded testing.
Knit Health’s NEA-led Series A funds deployment of Med-BEAM for care navigation and hospital operations. Its early company benchmark beats three Claude models at predicting clinical actions, while better patient outcomes remain unproven by that test.
Science Corporation’s Series D follows its European retinal-implant launch. Clinical results support improved central vision, but U.S. approval, eligible patients and commercial terms remain consequential questions for the startup, reportedly valued at $3.5 billion.
Google will let Gemini users save conversations while opting out of generative AI training on future chats. The global rollout also extends storage of history-off and temporary chats from 72 hours to around two months; human review and longer retention can still apply.
Google is removing the subscription requirement for Gemini’s task mode in six countries, while adding U.S. payment, calling and messaging features. The expansion broadens access to delegated work, but usage limits, permissions and website compatibility still govern what it can finish.
Eligible Pixel owners can join a Play Store beta to generate home-screen widgets from descriptions, but device, language and regional restrictions still apply.
Microsoft’s new in-house coding models compete with outside suppliers inside GitHub Copilot. Company tests show lower task costs at promotional prices, while its continuing financial ties to OpenAI complicate the rivalry.
Valkryn disclosed a $15 million seed round and early military testing contracts. Its evaluator-focused Battle Lab faces existing commercial tools, public testing capabilities and funding constraints that software alone cannot resolve.