Brief IA
›

Microsoft unveils seven models to cut agents' costs

Code & Dev·Tom Levy·

Microsoft unveils seven models to cut agents' costs

Microsoft unveils seven models to cut agents' costs
Key Takeaways
1Microsoft AI unveiled seven specialized models at Ignite.
2The promotional pricing for the code models runs until May 17, 2027.
3According to Microsoft, MAI Cyber already accounts for nearly 40% of MDASH analyses.
Why it matters : Microsoft offers lower-cost specialized models for continuously operating agents, whose token expenses are rapidly increasing according to the company.

Microsoft AI unveiled seven specialized models at Ignite, aimed at continuous AI agents. The company presents them as cost-effective alternatives to OpenAI and Anthropic systems. The comparisons are based on its own evaluations, with promotional rates for the code.

Microsoft Already Powers 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 mentions 10 million conversations each month and a halved error rate for multilingual processes. For its large in-house models, the company now uses its second-generation Maia 200 accelerator, which is said to produce 1.7 times more tokens per watt.

Microsoft introduces MAI Realtime 1 as its first speech recognition model. It is designed to account for interruptions, allow speakers to finish, and understand requests despite background 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 for Dynamics 365 Contact Center as well as in Dragon Copilot, where nurses need to simulate exchanges with patients.

Microsoft claims it achieves better results than GPT Realtime 2.1 on Big Bench Audio, τ-Voice, and trials with Dynamics 365, as well as the highest scores on Big Bench Audio for a relatively low cost. The service costs one cent per audio minute. Challenging tasks require a more powerful reasoning engine, charged at $1 per million incoming tokens and $5 per million outgoing tokens.

Microsoft Automates Treatment Selection

Microsoft explains that token spending is rapidly increasing among clients wishing to keep their agents active continuously. Presented as a block at Ignite, unlike the staggered releases seen at 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 that this routing achieves the accuracy of a large language model, with execution 2.2 times faster and a cost 4.4 times lower.

Decision also ensures classification, prioritization, and process control; in Cowork, its Auto function adjusts the reasoning effort. It accepts images, including screenshots, selects from actions permitted by developers, and returns calibrated confidence scores. Applications can request another treatment if uncertainty is high, while the model is designed not to decide when information is lacking. Google’s Gemma 4 26B A4B, a mixture of experts model, serves as its basis instead of Qwen3.5-9B; Microsoft has used hundreds of synthetic and licensed datasets for its supervised and reinforcement training before running it on its proprietary inference engine.

Decision scores 68 points on the Decision Index v2.1 from Artificial Analysis, six points higher than its predecessor, and ranks ahead of Jev 1.14, Gemini Decision 1, and GPT-6 Luna (Decisions). Microsoft also claims the best average accuracy among eight models across 102 internal evaluations totaling over 385,000 questions. The cost for one million incoming tokens is $0.042, or $0.05 in the European data zone; outputs are not charged.

In Foundry IQ, Microsoft-Retriever-1, available in a preliminary version, prepares searches within enterprise documents, examines their results, and completes them until sufficient elements are available to respond; Microsoft claims it is faster and more efficient than previous methods. The Foundry 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 utilize them in their Azure environment, exclusively through their Microsoft account team. Data and adaptations must remain with the client, and Microsoft commits to seeking explicit consent before any reuse for its training.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Code Pricing Includes a Promotion

In GitHub Copilot, one million tokens of MAI Code 2 costs $0.50 for input and $2.50 for output until May 17, 2027; prices are expected to rise to $1 and $5, respectively. For MAI Code 2 Flash, the corresponding rates are $0.05 and $0.25 during the promotion, then expected to increase to $0.10 and $0.50. Microsoft’s financial comparisons include a 50% discount on its model.

Microsoft positions 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 are based on DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1; three evaluations used its MSBench platform with GitHub Copilot CLI, without sub-agents. 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 business work and are not yet optimized for fundamental research or long-duration agent tasks.

The successor to MAI-Code-1.1-Flash, MAI Code 2 Flash targets explanations, precise edits, and ordinary corrections; Microsoft claims 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 operates within Microsoft 365 and Fabric IQ. According to Microsoft, it matches GPT-6 Luna in internal tests of Excel and PowerPoint, approaches GPT-6.1 Sol in Word with about 42% fewer incoming tokens, and surpasses GPT-5.6 Sol, GPT-6 Astra, and Opus 5.5 for DAX queries in Power BI.

Microsoft Excludes Tasks from CyberGym Testing

To evaluate MAI Cyber 2 Flash, Microsoft states it removed tasks from CyberGym that were common to some of its training data: 10% of the approximately 900 tasks relied on the same open-source projects and vulnerability reports. The company assures that it never integrated benchmark solutions into training and applied this exclusion to all compared models. On the retained scope of CyberGym-E2E, it reports 92.3% for MAI Cyber 2 Flash, compared to 91.5% for GPT-6 Astra, the latter utilizing twice the reasoning budget.

Microsoft estimates the cost per task at less than 2% of that of GPT-6 Astra, at $0.45 compared to $34.38 on CyberGym-E2E, with a price per token of one-hundredth. The displayed rates are $0.10 per million incoming tokens and $0.50 for output, compared to $10 and $50 for GPT-6 Astra. On WebExploitBench, Microsoft reports 40.4%, against 38.4% for GPT-5.6 Cyber. The comparison with GPT-6 Cyber, reserved for selected clients, is still pending: Microsoft cites the additional protections needed for agent evaluations using tools and anticipates results in the coming weeks.

The context of MAI Cyber 2 Flash reaches one million tokens, four times that of MAI Cyber 1.1 Flash. Outgoing tokens cost 86% less than those of its predecessor. According to Microsoft, MAI Cyber models already perform nearly 40% of the analyses on MDASH, its security platform.

MAI Image Ranks First in Retouching

Artificial Analysis ranks MAI Image 2.7 first for retouching, with 1,215 Elo points from 5,478 votes. It leads GPT Image 2.6 Sunburst (max) by four points for one-fifth of the cost but ranks fourth for images produced from text. Artificial Analysis estimates that versions 2.7 and 2.7 Flash offer retouching quality equivalent to GPT Image 2.6 Sunburst at lower rates. Prices are around $35 for 1,000 images with MAI Image 2.7 and $10 with Flash.

The model can modify the color, material, or location of an object while preserving other elements and produce transparent backgrounds. Smaller and priced at less than a third of the cost, the Flash version operates Canvas, the new editor from MAI Labs. MAI Image has been the default choice for Bing Image Creator since June. Microsoft also cites Adobe, Figma Weave, Minute (Sports Illustrated), Wirestock, and Manus among the companies working with its image models.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.