Microsoft AI presented seven specialized models at Ignite, aimed at continuously operating AI agents. The company offers them as lower-cost alternatives to OpenAI and Anthropic’s systems. The comparisons are based in particular on its own evaluations, with promotional prices for the code.
Microsoft already provides Dragon Copilot transcriptions
In Dragon Copilot, all transcriptions of exchanges between doctors and patients now go through Dragon Copilot Transcribe, based on MAI Transcribe 2. Microsoft reports 10 million conversations each month and an error rate halved for multilingual processes. For its large in-house models, the company now uses its second-generation Maia 200 accelerator, which is expected to produce 1.7 times more tokens per watt.
Microsoft presents MAI Realtime 1 as its first voice recognition model. This is designed to take interruptions into account, let people finish and understand requests despite the noise. Compatible with OpenAI’s Realtime API, it facilitates the transition of existing applications to this model; Agoda and Reservations.ai are testing it. Its integration is planned in Dynamics 365 Contact Center as well as in Dragon Copilot, where nurses must be able to simulate exchanges with patients.
Microsoft gives it better results than GPT Realtime 2.1 on Big Bench Audio, τ-Voice, and tests with Dynamics 365, as well as top Big Bench Audio scores for relatively low cost. The service costs one cent per audio minute. Difficult tasks require a more powerful reasoning engine, charged at $1 per million incoming tokens and $5 per million outgoing tokens.
Microsoft automates the choice of treatments
Microsoft explains that spending on tokens is increasing rapidly among customers wanting to keep their agents active around the clock. Presented en bloc at Ignite, unlike the staggered releases practiced during Build in June, its range covers programming, cybersecurity, image, voice and process decisions. In Copilot, Microsoft-Decision-1.1 already determines whether a request should be assigned to Copilot or Cowork. Microsoft announces for this switch the precision of a large language model, with execution 2.2 times faster and a cost 4.4 times lower.
Decision also ensures classification, prioritization and control of processes; in Cowork, its Auto function regulates the reasoning effort. It accepts images, including screenshots, chooses from actions allowed by developers, and returns calibrated confidence scores. Applications may seek alternative processing if uncertainty is high, while the model is designed not to decide when information is missing. Gemma 4 26B A4B, Google’s expert mix model, serves as its basis instead of Qwen3.5-9B; Microsoft used hundreds of synthetic and licensed datasets for its supervised and reinforcement training, before running it on its proprietary inference engine.
Decision reached 68 points on the Artificial Analysis Decision Index v2.1, six more than its predecessor, and ranked ahead of Jev 1.14, Gemini Decision 1 and GPT-6 Luna (Decisions). Microsoft also claims the best average accuracy among eight models out of 102 internal evaluations totaling more than 385,000 questions. The million incoming tokens cost $0.042, or $0.05 in the European data zone; outings are not charged.
In Foundry IQ, Microsoft-Retriever-1, available in preview, prepares searches in corporate documents, examines their results and completes them until there are enough elements to answer; Microsoft says it’s faster and more efficient than previous methods. Foundry’s Model Router now automatically selects models from an expanded set. With MAI Licensed Weights, companies should also be able to adapt certain models to their data and exploit them in their Azure environment, exclusively through their Microsoft account team. Data and adaptations must remain with the customer, and Microsoft undertakes to request their explicit consent before any reuse for its training.
The evening letterFree · 7 p.m.
Do you code with AI?
Deciphered AI dev tools, agents and new features, every evening in 5 minutes. Free.
Included upon registration: our selection of the best AI guides & comparisons.
Choose your pace
Free · No spam · Unsubscribe in 1 click
The code prices include a promotion
In GitHub Copilot, MAI Code 2’s one million tokens cost $0.50 in and $2.50 in output until May 17, 2027; prices must then reach $1 and $5 respectively. For MAI Code 2 Flash, the corresponding rates are $0.05 and $0.25 during the promotion, then need to increase to $0.10 and $0.50. Microsoft’s financial comparisons include a 50% discount on its model.
Microsoft rates MAI Code 2 at 98% of the performance of GPT-6.1 Sol for 40% of its cost, and at 96% of that of Claude Opus 5.5 for 5% of the cost. These averages cover DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1; three evaluations used its MSBench platform with GitHub Copilot CLI, without subagents. For interfaces, games, data visualizations and SVG graphics, the company also reports a 64% human preference over GPT-6 Sol in blind comparisons. However, these models are aimed at routine work in companies, without yet being optimized for fundamental research or long-term agentic tasks.
Successor to MAI-Code-1.1-Flash, MAI Code 2 Flash target explanations, precise edits and ordinary corrections; Microsoft says it is comparable to GPT-6 Luna for 65% of its cost and to Claude Haiku 5.5 for 20% of its cost. Its derivative MAI Work Flash works in Microsoft 365 and Fabric IQ. According to Microsoft, it matches GPT-6 Luna in Excel and PowerPoint internal tests, approaches GPT-6.1 Sol in Word with approximately 42% fewer incoming tokens, and outperforms GPT-5.6 Sol, GPT-6 Astra, and Opus 5.5 for Power BI DAX queries.
Microsoft excludes tasks from CyberGym test
To evaluate MAI Cyber 2 Flash, Microsoft says it removed common CyberGym tasks from some of its training data: 10% of the roughly 900 tasks relied on the same open source projects and vulnerability reports. The company assures that it has never integrated the reference solutions into the drive and has applied this exclusion to all the models compared. On the retained scope of CyberGym-E2E, it announces 92.3% for MAI Cyber 2 Flash, compared to 91.5% for GPT-6 Astra, the latter mobilizing twice as much reasoning budget.
Microsoft figures the expenditure per task at less than 2% of that of GPT-6 Astra, or $0.45 compared to $34.38 on CyberGym-E2E, with a price per token of one hundredth. The prices displayed are $0.10 per million tokens in and $0.50 out, compared to $10 and $50 for GPT-6 Astra. On WebExploitBench, Microsoft announces 40.4%, compared to 38.4% for GPT-5.6 Cyber. The comparison with GPT-6 Cyber, reserved for selected customers, is still missing: Microsoft invokes the additional protections necessary for agentic evaluations using tools and expects results in the following weeks.
The context of MAI Cyber 2 Flash reaches one million tokens, four times that of MAI Cyber 1.1 Flash. The exiting tokens cost 86% less than those of the predecessor. According to Microsoft, MAI Cyber models already perform nearly 40% of the analyzes of MDASH, its security platform.
MAY Image arrives first in retouching
Artificial Analysis ranks MAI Image 2.7 in the lead for editing, with 1,215 Elo points from 5,478 votes. It precedes GPT Image 2.6 Sunburst (max) by four points at a fifth of the cost, but ranks fourth for images produced from text. Artificial Analysis estimates that versions 2.7 and 2.7 Flash offer editing quality equivalent to GPT Image 2.6 Sunburst at lower prices. Prices are around $35 for 1,000 images with MAI Image 2.7 and $10 with Flash.
The model knows how to change the color, material, or location of an object while preserving other elements, and produce transparent backgrounds. Smaller and less than a third of the price, the Flash version runs Canvas, the new editor from MAI Labs. MAY Image has been the default choice for Bing Image Creator since June. Microsoft also lists Adobe, Figma Weave, Minute (Sports Illustrated), Wirestock and Manus among companies working with its image models.

