Today we’re announcing MAI-Cyber-1-Flash inside of MDASH, our multi-agent vulnerability identification and remediation harness. Together they deliver world-class performance at 50% of the cost of leading models.

Progress in AI has been startling and so has the new generation of cyber threats it’s unleashing. Attackers now wield increasingly powerful capabilities, probing an ever-growing mountain of code for just a single weakness that lets them in.

As the cost of finding a flaw collapses, the old model of security, where you scan occasionally and patch eventually, is now obsolete. If we’re to unlock the true benefits of AI, we must first build outstanding cyber models that help all of us harden the software the world runs on.

That’s the motivation behind MAI-Cyber-1-Flash, which has been built to find challenging vulnerabilities in complex codebases. It’s been deeply integrated into MDASH, honed by the best cybersecurity experts in the industry and hardened across the largest security estate on the planet.

This combined expertise delivers exceptional security protection, beating Mythos, Gemini and GPT on CyberGym, the gold standard benchmark for evaluating how systems reason over large codebases to find real vulnerabilities in the code.

Picking the right model for the task

Security is an always-on mission, and given the enormous volume of inbound attacks, token cost is now the real constraint for defenders. MAI-Cyber-1-Flash was designed to efficiently handle up to 90% of all tasks, enabling MDASH to use the larger and most costly models in our fleet (in this case GPT-5.4) for the 10% of exceptionally hard tasks that truly need them.

The result is that the unified system of MDASH with MAI-Cyber-1-Flash delivers 96% on CyberGym (on the benchmark’s any crash score; outperforms Mythos on the CyberGym leaderboard).

This combination delivers a 50% cost saving when compared against our best offering in MDASH today (GPT 5.4 + 5.4 mini + 5.3 codex). That’s the power of a well-tuned, multi-model system with access to uniquely rich historical training data. It ensures you always have the best model at the best price for every task.

In this new environment, being able to go from identifying a new vulnerability to addressing it in real-time is critical. And while AI remediation of software vulnerabilities is now a key security workflow, there are many jobs to be done by Security practitioners themselves.

That’s why today we’re also launching Perception, our agentic security systems, that provides teams of agents for a variety of security workflows in MDASH, to continuously monitor, patch, and close new threat vectors. Perception will also soon use MAI-Cyber-1-Flash for many more security workflows, beyond the software vulnerability work.

Three things matter today: Model. Data. Harness.

We have jointly optimized our world-class models, our unmatched historic data, and our expert-tuned harness to ensure that our customers have a uniquely powerful security offering.

Model. MAI-Cyber-1-Flash is a compact, code-heavy security model derived from the MAI-Thinking-1 lineage, which was built from scratch, in-house, on the highest quality data. Details in our technical report.

Data. Our deepest advantage. Decades of building world-class security systems now give us trillions of daily signals across identity, endpoint, cloud, and network, and an unmatched record of real exploits and remediations. No one can manufacture this history.

Harness. MDASH, our multi-agent vulnerability identification and remediation harness, is tuned by the best security experts in the industry, who have created 100+ agents using multiple leading models to find, validate, and remediate vulnerabilities. Agentic code scanning is a critical function in the Security Operating Center and feeds Project Perception, our new agentic security system.

Built with safety first

Because MAI-Cyber-1-Flash is Microsoft’s first cyber model, we built trust into every layer of the system, from model training to customer deployment. The model was developed with a security-first calibration, rigorously evaluated by Microsoft’s AI Red Team, tested through automated and expert-led adversarial exercises, and independently assessed by a third party.

Trust extends beyond the model itself. Through MDASH, customers get enterprise-grade controls including Role-Based Controls, tenant isolation, encryption, auditability, and sandboxed execution environments with no internet access. The result is a cyber model that delivers powerful capabilities to defenders while maintaining the governance, security, and control enterprises expect from Microsoft.

Our hill-climbing machine

Cybersecurity is not just a data-rich domain; it is a live reinforcement learning loop. Every day, defenders investigate threats, triage alerts, hunt adversaries, remediate vulnerabilities, deploy protections, and learn from the outcome.

Microsoft sees that loop end to end: vulnerabilities through Microsoft Security Response Center; attacks and defenses across identity, endpoint, cloud, data, browser, and applications; more than 100 trillion security signals every day; and operational insight from 1.6 million customers. Because we can connect actions to outcomes; what was exploitable, what was contained, what was blocked, and what actually worked; we have more than data.

Our MAI reinforcement learning loop gives us the foundation to build cyber models that improve continuously and become expert cyber defenders. That’ll remain our commitment to our customers for years to come.

Today we are introducing MAI-Thinking-1, Microsoft AI’s reasoning model. It is a medium-sized model that stands among the strongest models in its weight class. It matches leading models on key software engineering benchmarks, demonstrates advanced mathematical reasoning capabilities, and is preferred to Sonnet 4.6 in our blind human side-by-side evaluations. We don’t distill from other labs and we don’t rely on opaque data. Our datasets are clean, traceable, and enterprise-grade.

MAI-Thinking-1 is a step in our broader work to build towards Humanist Superintelligence: advanced AI capabilities designed to serve people and organizations, not to replace them. The model matters on both axes: what it can do, and how it was built.

The Hill-Climbing Machine

More than a single model, we are excited to introduce our Hill-Climbing Machine: a co-designed pipeline built to make every component of model development climbable, so capabilities improve continually and reliably over time. The aim is a repeatable system that can absorb better data, stronger rewards, more capable environments, and more compute.

Three main pillars guide our philosophy.

First, capabilities should be learned, not inherited. Although faster to acquire, inherited intelligence lacks the steerability essential for real world usage: an imitator is fundamentally tied to the design choices of its teacher and struggles to adapt to new situations. MAI-Thinking-1 was trained without distillation from third party models, forcing our model to truly learn the tasks at hand.

Second, clean data. We trained it from the ground up on clean, traceable and enterprise-grade data, without distillation from third-party models. This matters for quality, provenance, and control. If we cannot account for what shaped a model, we cannot fully understand its behavior or credibly improve it.

Third, self-sufficiency across the entire stack. All the way from co-design of our models with MSFT’s own accelerators through to our reinforcement learning framework, we have focused efforts on in-house training infrastructure. This is a crucial part of building our hill-climbing machine, to ensure we can fully optimize and shape our systems end-to-end to best serve our needs.

Medium-sized model, with strong software engineering performance

MAI-Thinking-1 is a 35B-active, ~1T-total parameters, sparse Mixture of Experts model, a smaller inference footprint than much larger models. Despite this, our model is toe-to-toe with Claude Opus 4.6 on SWE-Bench Pro. That matters for developers and enterprises because model size determines where advanced coding assistance can be deployed, how often it can be used, and whether it can move from exceptional tasks into daily workflows.

We have invested heavily in the training environments needed for agentic coding. Each verified environment is deterministic, executable, and graded by real test suites. This gives the model practice on the kind of multi-step work developers actually do: reading code, editing files, running tests, observing failures, and recovering from intermediate mistakes.

Advanced mathematical reasoning capabilities

MAI-Thinking-1 reaches 97.0% on AIME 2025, and 94.5% on AIME 2026, showing strong mathematical and scientific reasoning for its weight class. Strong performance here gives us confidence that our training loop can create real reasoning gains – climbing all the way from the ground up – from our own data, rewards, and evaluation process, enabling this intelligence to generalize to other domains over time.

Preferred in human side-by-sides vs. Sonnet 4.6

People care about whether a model understands the task, follows instructions, uses the right level of detail, writes clearly, and respects their time.

We built a blind side by side human evaluation with one of our partners, Surge, using their pool of professional raters to measure various models on these traits. The evaluation spanned 1,276 tasks across a wide variety of use cases in both single-turn and multi-turn conversations, with a focus on measuring how helpful each response is and whether it actually advances the user’s goals. In these evaluations, users preferred MAI-Thinking-1 over Claude Sonnet 4.6.

This has been a core focus of post-training. We want the model to be capable without being brittle, concise without being incomplete, and helpful without overreaching. Human preference data gives us a direct signal on whether benchmark improvements translate into better experiences for users.

Enterprise ready

MAI-Thinking-1 is built with enterprise readiness in mind. It supports long context with a 256k token window (enough to fit a 600 page document), function calling, and the flexibility to add developer instructions. We trained the model to follow multiple layers of instructions and aligned its default style to enterprise needs. It’s compatible with the widely used Chat Completions API. All MAI models come with enterprise-grade security and compliance through Microsoft Foundry.

Results

We report results in two views: post-trained MAI-Thinking-1 evaluations, and pre-training metrics for our base model.

Table 1. MAI-Thinking-1 metrics

Putting humans first

We are building towards Humanist Superintelligence: advanced AI capabilities designed to serve people and organizations, not replace them. Our models must remain subordinate technologies under human control with the goal of upholding human autonomy and being helpful. That means our models must not refuse legitimate requests under the guise of safety and compliance as then they are not truly serving humans.

Striking the delicate balance between being helpful and safe is not easy. For MAI-Thinking-1, we aimed to achieve this balance by treating unsafe compliance and unnecessary refusal as defects in the same reward construction where aggregation is based on severity of potential of harm. Safety is trained with the same reinforcement learning infrastructure used for capability, so safety rewards are part of the same hill-climbing loop ensuring safety is always aligned to the capabilities and not incidental.

As a result, we see that our model can balance ensuring a safety bar on sensitive unsafe requests while also being helpful on non-sensitive content.

Availability and access

MAI-Thinking-1 is available in public preview on Microsoft Foundry.

Try in Playground

MAI-Code-1.1-Flash produces higher quality code, at 25% greater token efficiency, and at a quarter of the cost compared to the model we launched in June at Microsoft Build. This small, efficient, coding workhorse is now in production in GitHub Copilot.

We learned from developer feedback that CLI tasks and .NET performance mattered, so that’s where we focused. The result: a 22% improvement on Terminal-Bench 2.1 in GitHub Copilot CLI and a 15% improvement on .NET tasks.

Benchmarks are useful guides but production is where the rubber meets the road. Most importantly, code survival rose 4% and return visits increased 9%.

1.1 is also dramatically more efficient. In GitHub Copilot tokens stream 25% faster and the model uses 25% fewer tokens to complete a task. That means faster answers, less waiting, and more useful work from every token—not simply a bigger model with a bigger bill.

Better training and serving efficiency let us offer a stronger, faster model at one quarter of the price of 1.0—and pass those savings reliably to customers. We achieved this by optimizing for real-world use across more than hundreds of thousands of reinforcement-learning environments in GitHub Copilot.

The loop is simple: ship, learn, improve, repeat. That’s the MAI hill climbing machine.

Help shape future improvements

Try MAI-Code-1.1-Flash today in GitHub Copilot, then tell us what needs improving by opening an issue here.

Today, we’re announcing MAI-Image-2.6 – ranked second on the Arena text-to-image leaderboard.

It’s another significant climb for the MAI-Image family, improving +79 Elo over MAI-Image-2.5 overall, with gains across every Arena text-to-image category. Text rendering alone improves by +91 Elo.

This result firmly establishes MAI-Image ahead of leading models from Meta, Google and xAI.

Another step up in quality

With every MAI-Image release, we’re focused on moving the quality frontier forward.

MAI-Image-1 gave us the foundation. MAI-Image-2 made a major jump in photorealism, text and creative range. MAI-Image-2.5 pushed further into professional-grade imagery and editing.

MAI-Image-2.6 continues that climb with broad gains across different categories that we know our users care about most:

• Stronger text rendering
• Better portraits and 3D imagery
• More polished commercial and photorealistic outputs, with stronger results across product, branding and cinematic use cases

In the Arena evaluations, MAI-Image-2.6 significantly improves on 2.5 in every measured category.

And there’s more to 2.6 – from working across multiple references and richer grounding to greater control over reasoning, format and resolution. We’ll share more on that soon.

More to come

MAI-Image-2.6 is another step in building our hill-climbing machine for image generation: continuously improving quality, expanding capability, and turning those gains into models that are more useful for real creative work.

Try MAI-Image-2.6 for text-to-image today on Arena. Coming later this week to MAI Playground, and rolling out soon across Microsoft Foundry and other products.

Tokenmaxxing has been the story of the last few months, but token efficiency is the next big focus across the industry. How do we get the best possible performance per token invested, and the best real customer outcome per dollar invested?

To build a frontier firm, you have to optimize frontier performance against cost. Choosing where you want to sit on that curve is critical. By co-optimizing your models, harnesses, and RLEs you can pick a point on the curve that suits your firm.

In most cases, frontier generalist models aren’t necessary for every task. By tuning models for a specific product, you can maintain or even exceed frontier performance, while reducing token costs dramatically.

This is where we have focused our MAI hill-climbing machine over the last quarter, and the results are pretty cool. This week we released MAI-Cyber-1-Flash optimized for our MDASH harness.

Together, the system landed at No.1 on the leading CyberGym benchmark – beating Mythos by 12ppts – at 50% of the cost. And remarkably, we serve it on H100s too.

It was designed to handle up to 90% of tasks efficiently, so that MDASH can reserve the largest and most expensive models in our fleet (in this case GPT 5.4) for the 10% of exceptionally hard problems that truly need them.

As Satya mentioned today in our Q4 Earnings call, since last quarter, we’ve shipped more than a dozen new models across image, voice, transcription, coding and security, and they’re already powering many of Microsoft’s most widely used products to maintain or improve quality while using significantly fewer tokens, in many cases saving 50-90% of GPU costs:

• We built MAI-Code-1-Flash hand-in-hand with our colleagues at GitHub, where since June millions of developers have used it in their daily work. 10% higher code accept rate and 10% lower median token usage than GPT-5.4 Mini and Claude Haiku 4.5 in VS Code, and already showing improved retention.
• We then trained that same checkpoint inside an Excel RL environment to achieve comparable performance to GPT-5.6 for the most common tasks while being more cost-efficient, and small enough to serve on an A100 or H100 vs only the latest and most expensive accelerators.
• MAI-Image-2.5-Flash, is now the end-to-end default in Bing Image Creator, in production in PowerPoint where it is reducing GPU costs up to 84% compared with GPT-Image-2, and is the default for key OneDrive editing scenarios, where it has increased save rates by 26% and delivers up to 2.5x greater token efficiency.
• MAI-Voice-2-Flash now powers Dynamics 365 Contact Center, where customers like T-Mobile and EasyJet build their call center agents, reducing GPU costs by up to 89%.
• MAI-Transcribe-1.5 now serves Dragon Copilot’s multilingual workflow across 58 languages — a solution used by 170,000 medical providers that processed 28 million patient encounters last quarter, where our tests show a 50% relative in reduction transcription and language-identification error rates.

And what’s more, by co-designing our models with our own silicon, we are seeing 40% better performance per watt running MAI models on Maia 200.

But the benefit is not only cost. It’s resilience. Every business now must assume that any one model it depends on could disappear, through a security incident, a business or policy misalignment, or a geopolitical shift.

Every model in a product or agentic system should be substitutable, and that’s only possible when you build the harness, context, memory and action space independently of a single model family. That’s the hill-climbing machine we’ve built.

We think this is the beginning of a genuinely new performance curve. Its shape represents a system rather than a model, and traversing this curve delivers better quality, lower cost, and more choice.

This has been a summer of hard but wonderful work by the team. We are keenly aware of how early this is, and of how much we still have to learn. But the direction is clear, we are hill-climbing to move the frontier on the cost-to-outcome curve, and we will keep sharing what we learn along the way. There is much more to come.

Better models, fewer parameters, less tokens

At Build in June, we introduced our hill–climbing machine, our integrated data, model, and harness flywheel. Today we are excited to share two examples inside Microsoft: MAI models specialized for agentic workloads in GitHub Copilot and Excel.

Early results are promising. In our live product deployment, we see that our MAI model deployed in Excel is on par with GPT-5.6 for the most common tasks while being more cost-efficient.

The figure below shows how MAI-Code-1-Flash, post-trained within the GitHub Copilot harness, was used as the starting checkpoint to climb on Excel evaluations, resulting in two highly efficient, specialized models.

MAI-Code-1-Flash in GitHub Copilot

Since launching MAI-Code-1-Flash in GitHub Copilot in June, millions of developers have been using it for their day-to-day work, where it’s outperforming other similarly sized models while using fewer tokens.

• It has an approximately 10% higher code accept rate than GPT 5.4 Mini and Claude Haiku 4.5 in VS Code.

• Developers were 6% more likely to return across multiple days than with GPT 5.4 Mini and 11% more likely than with Claude Haiku 4.5.

• It has 10% lower median token usage than GPT-5.4 mini and Claude Haiku 4.5, with more user-initiated turns.

MAI model live in Excel

Excel offered a test for whether the capabilities built into MAI-Code-1-Flash could transfer beyond the domain they were trained for, moving from agentic coding to agentic knowledge work. To do so, we further trained our MAI-Code-1-Flash checkpoint in an Excel reinforcement learning environment to learn about tools and knowledge workflows in spreadsheets. The result is a model with a command of Excel workflows that is more efficient and less expensive to run.

User feedback from production traffic indicates that the quality of the MAI model in Excel is on par with GPT-5.6 for the most common tasks. In addition to the direct model cost savings, this smaller, more efficient model can be served on both Nvidia H100 and A100 class GPUs rather than requiring only the latest-generation accelerators, which significantly lowers the cost of deployment for Microsoft.

Training agentic models from inside the product stack

These results point toward a broader strategy. By having access to the entire product stack—the model, the harness that runs it, the agents, and product-specific evaluations—we can hill-climb to train efficient, powerful models capable of tasks previously handled by larger, more expensive ones.

Beyond GitHub Copilot and Excel, we’re currently extending this hill-climbing approach to train efficient models across Microsoft’s family of agentic products: Copilot Chat, Outlook, PowerPoint, and more.

A year ago, we set out to develop purpose-built models in-house at Microsoft AI. Models trained on clean, traceable, enterprise-grade data, without distillation from third-party models, and designed from the ground up to serve the people who use Microsoft products every day.

Today, that work is showing up where it matters in the products you already rely on. The models we previewed at Build are not just topping leaderboards, they are running in production, at scale, efficiently powering experiences for millions of users across a growing portfolio of Microsoft products, including Bing, PowerPoint, OneDrive, Dynamics 365, and Azure.

Introducing two new model variants

Real-world use cases are not one-size-fits-all. A creative studio chasing maximum fidelity has very different needs from a customer service center serving millions of calls daily. That’s why we’re building families of models: to give every product the right balance of quality, speed, and cost.

Today we’re expanding those options:

• MAI-Image-2.5-Pro is now in public preview. For use cases where quality is top of mind. Hero imagery, detailed editing, precise in-image text rendering. Pro is our highest-fidelity image model to date, priced at $5 per 1M text input tokens, $8 per 1M image input tokens, and $106 per 1M image output tokens.
• MAI-Voice-2-Flash is now in public preview. First introduced at Build, Flash is built for speed and scale. It’s the fast, efficient path for high-volume voice experiences where responsiveness is everything, all while retaining the natural prosody and high acoustic quality found in MAI-Voice-2. Flash is 2x faster than MAI-Voice-2 and 32% cheaper, priced at $15 per 1M characters.

MAI-Voice-2-Flash and MAI-Image-2.5-Pro sit alongside our existing production models so builders can pick the point on the quality-speed-cost curve that fits their job.

Purpose-built models for our Microsoft products

Leaderboards are great for benchmarking the quality of our models, but the proof lies in serving users in real product use cases. Microsoft rigorously evaluates all models before deciding on the best one for use in production environments. MAI models are outperforming competitive models for quality, latency and efficiency for more of Microsoft’s product surfaces:

• Bing Image Creator is now 100% in-house by default. With MAI-Image-2.5‘s new precise image editing capabilities, it is now the default model powering Bing Image Creator end-to-end, for high-quality generation and greater creative control.

Each of these enhancements is a step toward the same goal: Microsoft products, powered by Microsoft models, built to serve the people who use them.

Built with the experts

Some of today’s most consequential and challenging fields, such as medicine or software engineering, demand more than a general-purpose model. They demand deep expertise, tight feedback loops, and models that are capable of solving complex, real-world challenges. Microsoft AI partners directly with domain specialists to adapt and refine our models for their unique needs:

• MAI is partnering with Dragon Copilot, a solution used by 170,000 medical providers, which processed 28 million patient encounters last quarter.MAI-Transcribe-1.5 now supports Dragon Copilot’s multilingual workflow, replacing the previous model with our own best-in-class model across 58 languages. In internal evaluations on multilingual recordings, the model delivers a 50% relative reduction in both transcription and language-identification error rates across most languages, and early research shows promising improvements in the downstream accuracy of medical notes.

None of this is an endpoint. It’s the compounding result of one decision: build models in-house so they can be shaped around the people who use our products, and offer options that deliver customers more choice and better value. That’s what “powered by MAI” means, model by model, product by product. We’re just getting started. In the meantime, get started with the models in Foundry. You can also learn more about MAI-Image-2.5-Pro and MAI-Voice-2-Flash on Microsoft Learn or try them out in the MAI Playground.

William Sieghart is a world-renowned patron of poetry, a best-selling author of the “Poetry Pharmacy” anthologies who helped establish National Poetry Day in the UK. But it is at his one-on-one “consultations” at events across the country where he brings the power of poetry to life.

During these sessions, participants describe what’s on their mind and William recommends a poem. His measured, warm and compassionate approach has resonated with thousands of people for whom he has prescribed poems.

William asked Microsoft AI to help him scale these sessions to a global audience. Ode Poetry, the result of this collaboration along with creative studio Gravity Road, leverages MAI’s state-of-the-art audio models to deliver a first-of-its-kind digital poetry experience.

“When William approached us with the concept for Ode Poetry, it felt like the perfect opportunity to bring our MAI audio models to life in a way that reflects our humanist-first values,” said Trevor Back, MAI product lead.

Ode Poetry asks you questions about how you’re feeling and recommends a poem that might resonate, playing a recording of a real person reading it. AI simply supports the connection between the listener, the creativity of the original poem, and its human reader. It does this through empathetic questions and a thoughtful approach to recommending poems based on William’s literary analysis.

The project presented an interesting technical challenge. In order to build a voice experience that responds empathetically to users in real time and recommends an appropriate poem, the system must:

• Listen carefully and understand any user
• Respond at the right time, in the right style
• Remain grounded in William’s unique style and poetry expertise

We used MAI’s voice models to power this experience, and worked closely with Ode to build custom solutions for their specific needs, including safety and response quality, all detailed below.

Speech understanding with MAI-Transcribe

Ode Poetry uses our MAI-Transcribe model to understand spoken language in real time. A leader in its class, it can handle the nuances and complexities of verbal communication, including the many languages, accents, dialects and styles people use today.

The data reflects this. The latest version of MAI-Transcribe achieves the lowest Word Error Rate (4.86%) of any competitor on the FLEURS benchmark, showcasing its ability to understand many languages and difficult audio examples (such as background noise). The model enables users to engage naturally rather than adapting their speech to the technology. Real-world use cases such as Ode Poetry demonstrate the model’s ability to work effectively in production environments.

Bringing expressive speech to conversational AI with MAI-Voice

Ode Poetry also uses MAI-Voice to re-create William’s distinctive vocal style. What makes MAI-Voice special is its ability to reproduce William’s speaking style with very high fidelity, using only a short sample of audio. Once you’ve heard William’s voice, you’ll understand why this is so important.

MAI-Voice’s capabilities were achieved by building our own large-scale multilingual data pipeline, augmented with high-fidelity studio recordings of voice talent. This enabled researchers to train a model with both the breadth to sound natural across contexts and the depth to faithfully reproduce any specific speaker. Here’s an example:

An audio sample showcasing the natural expressivness available with MAI-Voice.

Reasoning between conversation and poem recommendations

Ode Poetry is a unique conversational experience because it not only enables a natural conversation, but it also recommends a poem as an outcome, using the structured conditions William has provided for the poems in his “Poetry Pharmacy” books.

We developed a harness to distill William’s years of experience with personal consultations into a system that can match the unstructured conversation with the user into a framework that enables the right poem to be recommended. This system included carefully chosen deterministic logic and tool-calling mechanisms.

We also focused on responsible AI for a safe user experience; for example, experimenting with jailbreaks that can trick a language model into attempting real psychiatric diagnoses. Through an iterative approach of safety evals, red teaming, and guardrail updates, we were able to prevent a range of misaligned model behaviors.

Hill-climbing was also a critical component to achieving a premium editorial product. High-quality production AI systems hinge on countless small improvements. Accuracy, consistency, latency, reasoning, and expressiveness all contribute to the final result, and a fix in one dimension can cause regressions in another. We evaluated Ode Poetry across a wide range of emotional scenarios and used the results to iteratively refine prompts, recommendation logic and speech output. Achieving a consistently excellent result took relentless iteration of both text and audio prompts, and the shared commitment from both Microsoft AI and William’s team made this collaboration a joint act of creative care.

This is reflected in the app launching today. As William shared when he first began interacting with his digital extension: “One of [the researchers] recorded my introduction and the next thing I knew, he was playing my voice…and the intonation, the pauses…I just thought, wow, that’s quite nice.

“It’s really exciting for me,” he added, “because what we’re about to embark on together is in a way my lifetime’s ambition, which is to bring poetry to everyone.”

We invite you to try out the experience of a poetry session with William at Ode Poetry, and test the latest versions of our models at the MAI Playground.

Today we’re launching MAI-Voice-2 — the most expressive, natural-sounding text-to-speech model we’ve built to date. It’s a significant leap from its predecessor across every dimension that matters to production voice experiences: fidelity, language coverage, speaker consistency, and emotional range. It is built for the products and services where voice quality directly impacts user experience: assistants or customer support that represent your brand, audiobooks that hold attention over hours, and accessibility experiences where voice is the only interface. It’s also built with responsible deployment in mind, with consent guardrails ensuring the technology is as trustworthy as it sounds. MAI-Voice-2 is now available in Microsoft Foundry, and is being integrated into VSCode and the Dynamics 365 Contact Center.

Features and capabilities

• Expanding from English‑only to 15 languages while maintaining the same naturalness and expressiveness as English.
• Granular emotion control via emotion tags: sad, whispered, excited, etc.
• Zero-shot voice prompting using 5-60s of reference audio available for all supported languages, with built-in consent guardrails.
• MAI-Voice-2 is preferred over its predecessor MAI-Voice-1 72% of the time.
• Stable speaker identity across long-form content – audiobooks, podcasts, lectures.
• Code-switching capabilities for select language pairs — such as Hindi-English and Spanish-English — matching the way users naturally mix languages in everyday speech.

Hear it for yourself:

English (emotion: Embarrassed)

German (emotion: Confused)

Hindi (emotion: Excited)

English (role: Motivational Trainer)

English (role: Sports Commentator)

Performance

MAI-Voice-2 generates very natural speech in a controllable way. In side-by-side preference tests, it was preferred over its predecessors 72% of the time. In speaker similarity evaluations, speech generated by MAI-Voice-2 is indistinguishable from recordings of the same voice. Below, you can verify this yourself by trying to identify where the human speech ends and the MAI-Voice-2 output begins.

Guess the human recording vs. MAI‑Voice‑2

Listen to the audio clips below – each blends human recordings with speech generated by MAI‑Voice‑2. Can you tell where the human voice ends and the synthetic voice begins, or vice versa? Or does it sound like one continuous voice?

Human recorded + TTS

Language: English US

Human recorded + TTS

Language: Hindi (India)

Human recorded + TTS

Language: Spanish (Mexico)

TTS + Human recorded

Language: French (France)

Human recorded + TTS

Language: German (Germany)

Supported Languages

We prioritized depth across 15 languages, ensuring for supported languages we support a spectrum of expressive capabilities spanning tonal, pitch accent, stress timed, and syllable timed systems. We plan to continue expanding and refining the expressive range for all supported languages.

MAI-Voice-2 now supports the following languages/locales: English (US), English (Australia), Italian, French, German, Hindi, Spanish (Spain), Spanish (Mexico), Portuguese (Brazil), Portuguese (Portugal), Korean, Chinese (Simplified), Turkish, Russian, Thai, Dutch, Romanian and Hungarian.

In markets where people naturally mix languages, we support code-switching – notably Hindi–English and Spanish–English – reflecting how people actually speak. In internal testing, the model switches languages mid sentence fluidly, without losing prosodic naturalness nor speaker identity.

Hindi + English

Spanish (Mexican) + English

Voice Synthesis

Developers can create a custom voice in Microsoft Foundry across all supported languages using just a short reference clip – no retraining or fine tuning required. With only a few seconds of audio (recommended: 5–60 seconds), MAI Voice 2 can generate high quality speech that matches the speaker’s identity, making it easy for companies to bring their own brand voice into products without maintaining a separate voice model.

Consent and Safety

Consent is enforced at the system level: only authorized, licensed voices can be synthesized in production. No unlicensed voice cloning is possible. To gain access to this feature apply here.

Use Cases

• Assistants: Branded voices for Copilot, apps, devices, customer support.
• Entertainment: Characters for games, podcasts, audiobooks, AR/VR.
• Accessibility: Narration for visually impaired users; voice for speech impairments.
• Education: Instructors and characters for courses and simulations.
• Creators: Turn text into audio with your own voice. No studio required.

Try it out

DuoAI

DuoAI is an experimental experience that gives you a direct way to try MAI‑Voice‑2, MAI‑Transcribe‑1.5, and MAI‑Image‑2.5 models in action – showcasing natural, fluid, expressive dialogue. In the demo, you can engage in a three‑way conversation with two agents and even generate images using MAI‑Image‑2.5. It’s a practical preview of how MAI multimodal models work together to build powerful, customizable voice agents. Try DuoAI now

Note: DuoAI is not meant to showcase the capabilities of the underlying LLM – that component is modular and can be swapped as needed.

You can also explore the models directly in the MAI Playground.

Learn more about MAI-Voice-2

• Model card [Link]
• Foundry API documentation [Link]
• Cookbook [Link]

MAI-Image-2.5 is our strongest image model yet – and now ranks No. 2 on Arena’s Image Edit leaderboard, ahead of Nano Banana 2.¹

Built for high-quality generation and precise, controllable editing, it brings production-ready image workflows to developers and Microsoft products.

Today, we’re launching MAI-Image-2.5 for maximum fidelity, and MAI-Image-2.5-Flash for fast, scalable production workloads.

Features and capabilities

Step-change in text-to-image quality

MAI-Image-2.5 produces more detailed, coherent images from prompts, with stronger text rendering, product imagery and prompt adherence.

Complex visual reasoning

The model understands scene structure, lighting, scale, and spatial relationships, helping it make edits that fit the image context, such as adding an object with the right perspective and shadows.

Fine-grained edit control

MAI-Image-2.5 supports precise, localized edits, from replacing an object or updating text to removing motion blur, without changing the rest of the image.

Face and identity consistency

MAI-Image-2.5 preserves facial identity across edits, maintaining recognizable likeness even through changes in pose, expression or viewpoint.

Benchmarks

MAI-Image-2.5 achieves Arena scores that surpass GPT-Image-1.5 and Nano Banana Pro 2K, ranking No. 3 for text-to-image and No. 2 on Arena’s image-editing leaderboard.

Across these evaluations, MAI-Image-2.5 demonstrates leading performance in image generation and editing, with strong results across prompt adherence, visual quality, and controlled image modification.

Powering Microsoft products

MAI-Image-2.5 is live on PowerPoint for high-quality image generation and rolling out to OneDrive for precise editing.

In PowerPoint, users can generate presentation-ready visuals and slides from prompts, turning ideas into polished decks faster.

In OneDrive, users can make precise photo edits – removing unwanted distractions, cleaning up backgrounds, and enhancing images while preserving the original scene.

Best price-to-performance models

MAI-Image-2.5 is available to developers in Foundry today, delivering premium quality and fine-grained editing control at $5 per 1M text input tokens, $8 per 1M image input tokens, and $47 per 1M image output tokens.

MAI-Image-2.5-Flash offers faster, lower-cost generation and editing at $1.75 per 1M text input tokens, $1.75 per 1M image input tokens, and $19.50 per 1M image output tokens.

Together, they give customers the flexibility to optimize production image workflows for fidelity, speed, or cost, while delivering leading price-to-performance on Arena score.

Safety and limitations

MAI-Image-2.5 includes layered safety guardrails, including prompt and output filtering, to help detect and block harmful or policy-violating content.

Like all image models, MAI-Image-2.5 can reflect biases in its training data and may produce plausible but inaccurate or misleading visual details. Generated images should be reviewed before use in sensitive contexts, including identity, legal, medical, financial, or news-related workflows.

Try it out

MAI-Image-2.5 and MAI-Image-2.5-Flash are now available to developers in Foundry, bringing high-quality image generation and precise, controllable editing to production workflows.

You can also try the models directly in the MAI Playground.

OpenRouter is also making MAI-Image-2.5 available to its developer community:

“We’re excited to bring Microsoft’s MAI models to OpenRouter. MAI-Image-2.5 is one of the strongest image models available today, and expands the set of multimodal capabilities available to developers on OpenRouter. Our goal is simple: when great new models launch, the 9 million developers building on OpenRouter should be able to use them immediately through the same API they already use.”
– Alex Atallah, CEO, OpenRouter