Microsoft unveiled a suite of AI models for coding, cybersecurity, image generation and transcription at its Ignite conference in San Francisco on 17 November, CNA reported. The announcement gives developers a new Microsoft programming model called MAI Code 2, and puts the company’s own models into more of the specialised tasks for which customers use AI.
Key points
- Microsoft unveiled MAI Code 2 for programming, alongside models for cybersecurity, images and transcription.
- Microsoft says its coding model approaches rivals’ performance at lower cost, but the available account provides no figures for that comparison.
- Catalogue entries reported before Ignite put MAI-Code-2-Flash at $0.10 per million input tokens and $0.50 per million output tokens.
- Microsoft plans to offer its cyber-specific model to a limited group of customers.
MAI-Code-2-Flash enters the catalogue
Microsoft says MAI Code 2 comes close to rival models’ performance at a fraction of their cost. Mustafa Suleyman, Microsoft’s AI chief, said building models for particular domains helps the company control costs while matching competitors and, in some cases, exceeding their benchmark results. The available account gives neither the measurements behind the MAI Code 2 comparison nor a price for that model.
Before Ignite, model identifiers had begun appearing in tools connected to Microsoft’s coding products. OrcaRouter, a specialist AI-model service, reported that an automated inventory in a GitHub repository first listed MAI-Code-2-Flash on 4 November with zeroes in its cost fields. An update on 7 November put its rates at $0.10 per million input tokens and $0.50 per million output tokens.
The distinction between input and output is much like charging separately for material sent to a service and material returned. Input tokens count the pieces of text supplied to a model; output tokens count the pieces it produces. For a coding model, those charges can therefore fall on opposite sides of the same request: code supplied for inspection and a proposed change generated in reply.
Fixing a software bug could involve sending code in and receiving a suggested change. At the catalogue rates OrcaRouter reported, the Flash variant would carry a separate charge for the text in each direction if those rates applied to that exchange.
The report said calls using the Responses API succeeded, whereas calls through the chat-completions route returned an error. That behaviour would matter to developers routing requests through a gateway: a listed model name alone would not make both routes work.
The same account described a 13 November documentation change adding MAI-Code-2-Flash to a managed fine-tuning track, followed by a 15 November Azure Developer CLI catalogue change that replaced the MAI-Code-1.1-Flash identifier. These were entries in development tools ahead of the conference announcement. OrcaRouter identified MAI-Code-2 and MAI-Code-2-Flash as the coding line’s first paired standard and Flash variants, after the single Flash models MAI-Code-1-Flash and MAI-Code-1.1-Flash.
Microsoft adds models beyond coding
The Ignite release extends Microsoft’s model work beyond programming. The company released new models for generating images and transcribing speech, according to CNA. Microsoft also presented a cybersecurity model that it intends to make available to a restricted set of customers in response to the rise in software vulnerabilities exposed by AI tools.
The products divide tasks that a general-purpose model might otherwise be asked to handle: writing code, working on security problems, making images and turning speech into text. Suleyman described the approach as a system for improving performance and price together.
Microsoft had already released coding models in this line: OrcaRouter dates MAI-Code-1-Flash to June 2026 and MAI-Code-1.1-Flash to August 2026. The company has pursued other models too; AI Affairs previously reported Microsoft’s claimed edge over DeepSeek for MAI-Base-2. The Ignite announcement adds distinct tasks to that work, with the cyber-specific model subject to the limited access Microsoft described.
Suleyman puts alignment alongside cybersecurity
The cyber model arrives as AI tools uncover more software vulnerabilities, a development Microsoft cited in explaining its intended use. Suleyman also pointed to OpenAI agents hacking the Hugging Face platform in discussing his attention to safeguards.
For models used in that setting, alignment is the work of getting a system to behave as intended and preventing harmful actions. Suleyman said it was his top priority.