Microsoft promises cheaper coding with MAI Code 2

At Ignite 2026, Microsoft unveiled three AI models, including MAI Code 2, which it claims achieves 96% of the performance of Claude Opus 5.5 for just 5% of the cost. These comparisons focus on API consumption, not on subscriptions. The company is integrating its models into various software, from GitHub Copilot to Microsoft 365.
Microsoft Implements a Launch Discount
The rates used for comparisons benefit from an initial 50% discount, Microsoft specifies. The choice of tests may also favor the displayed results: the representativeness of this selection remains uncertain. Therefore, an offer priced at $100 per month does not equate to a subscription of five dollars.
Additionally, Microsoft excludes from the targeted scope the most demanding research work and agentic tasks. Solving a mathematical problem that has been posed for decades is not among the likely applications of these models.
Microsoft Could Reduce Its External Spending
Since 2023, Microsoft has invested billions of dollars in OpenAI, whose intellectual property underpins many of the group's services and products. Until now, nearly all AI-related operations involved payments to providers of advanced models, borne by Microsoft or its clients. During his presentation at Ignite, Satya Nadella emphasized the now central role of AI in the company's activities. The slides showcased a wide range of foundational models intended for Copilot as well as other group offerings.
Microsoft is now evaluating its own models against those of OpenAI and Anthropic. To handle a large volume of ordinary requests, sufficiently reliable internal systems could limit reliance on these providers. The mentioned office operations would not necessarily require the highest capabilities. Less expensive models could thus yield savings of several hundred million, increased vertical integration, and greater control, without a direct intention to replace external offerings being established.
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
However, these savings are not reflected in observed financial results at Microsoft. A hypothetical scenario reduces the cost per request from $0.01 to $0.001, rather than to $0.0005, to eliminate the effect of the initial promotion. A company spending one million dollars per day on inference could then spend about $100,000 daily. This scenario suggests a potential annual saving of $325 million.
Microsoft Diversifies Coding in Its Software
For ordinary programming needs, Microsoft distinguishes MAI Code 2 from Code 2 Flash. The former is also presented by the company as comparable to GPT-6.1 Sol; the latter is likened to GPT-6 Luna and Claude Haiku 5.5, with a significantly lower announced cost. GitHub Copilot, Microsoft 365, and Fabric welcome these new models among their options.
The transition from coding to office work already has precedents: ChatGPT Work and Claude Cowork stem from agentic programming tools. Users had repurposed Claude Code and OpenAI Codex for productivity activities before AI companies offered more accessible work tools for non-programmers. In these various uses, agents must understand a complex environment, manipulate elements with tools, and provide reliable results. Because its outcomes are highly measurable, programming serves to develop and assess these capabilities.
Microsoft presents its office model MAI Work Flash as derived from Code 2 Flash. According to the company, it can assist in creating Excel formulas, work on Word documents and instructions, or track decisions and extract useful information in Teams. Reformatting a PowerPoint presentation is also among the likely uses.
Finally, Microsoft claims very good results for visualization tasks. These capabilities would be suitable for designing user interfaces as well as creating charts and slides.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.