Microsoft has unveiled a new family of in-house artificial intelligence models at its Ignite conference in San Francisco, pitching them as cheaper rivals to the best-known systems from OpenAI and Anthropic — and giving its Copilot assistant the ability to run long tasks on its own.
Chief executive Satya Nadella opened the conference on Tuesday with a keynote setting out what the company calls a "frontier ecosystem": its own models, its own chip, and a Copilot that behaves less like a chatbot and more like a colleague.
A cheaper coding model, built in-house
The headline announcement was MAI Code 2, a coding model Microsoft says is designed for the everyday work of software teams — taking code into production, deep debugging, multi-file changes and reasoning across a whole repository. A smaller sibling, Code 2 Flash, is aimed at high-volume routine jobs such as explaining code, making targeted edits and fixing bugs.
Both are available now in GitHub Copilot, in the model picker and inside Copilot Fusion, the system that chooses which model should handle each task. They are also available in Foundry, Microsoft's platform for building AI applications.
Microsoft claims Code 2 delivers 98 per cent of the coding quality of OpenAI's GPT-6.1 Sol at about 40 per cent of the price, and 96 per cent of Anthropic's Claude Opus 5.5 at roughly 5 per cent of the price. Those figures come from Microsoft's own tests, and should be treated with care: Axios noted that the comparisons rely on selected benchmarks and that some of the price advantage reflects promotional discounts.
The company also introduced MAI Cyber 2 Flash, a security model it says scored 92.3 per cent on a real-world vulnerability-finding test, slightly ahead of GPT-6 Astra, at under 2 per cent of the cost per task. It will reach customers through Microsoft's own security products from mid-December.
Copilot learns to work on its own
The other big theme was autonomy. Microsoft introduced Copilot Autopilot, which can carry out long-running, repetitive tasks and projects without step-by-step instructions, and Copilot Workspaces, where people and AI agents share the same context. A new "Today" experience prepares a user's next steps in advance, while Copilot Code — for building dashboards, trackers and apps through natural language — is due to reach general availability in December.
Microsoft also said trusted customers will now be able to access its model weights through a programme called MAI Licensed Weights, and it announced Maia 300, a new AI chip it says will deliver 40 per cent more tokens per watt than leading GPU-based systems.
The bigger picture
The announcements underline how far Microsoft has moved towards building its own AI, rather than relying solely on OpenAI, its long-time partner. Microsoft AI chief Mustafa Suleyman told Axios the company was making strong progress on its own frontier general-purpose model, though he declined to give a release date.
For developers and businesses, the pitch is simple: near-frontier performance at a much lower price, with the model choice handled automatically. Whether the savings hold up once the launch discounts fade is the question to watch — but for now, the AI coding market has a serious new competitor.
