When AI Gets Cheap Enough to Run All Day
Microsoft launched seven specialist AI models at Ignite that cost up to 95% less than the best models on the market — and that changes the economics for anyone building automated workflows.
- 7
- new AI models launched at Ignite
- 95%
- less than Claude Opus 5.5 for coding
- <2%
- cost of MAI Cyber vs. GPT-6 Astra
- 4,000
- businesses already using Decision 1.1
Microsoft launched seven new AI models at its Ignite conference on November 17. The most striking detail was not what they can do — it was what they cost. Its new coding model, MAI Code 2, costs 95% less than Claude Opus 5.5 while still delivering 96% of its performance. That gap is wide enough to change how companies think about building with AI.
The seven models — called the MAI family — each cover a different job: writing code, finding security holes, editing images, handling voice, and routing tasks to the right model. MAI Cyber 2 Flash, the security model, costs less than 2% of a comparable OpenAI model while matching its quality on a standard cybersecurity benchmark. These are not general-purpose models trying to do everything. They are focused tools built to run one job continuously and cheaply. Microsoft is also letting businesses download the model weights and run them on their own private servers — important for anyone in finance, healthcare, or law where data cannot leave the building.
This matters most for teams running AI agents — programs that do work automatically on your behalf, often around the clock. An agent checking code quality, monitoring for security problems, or processing customer requests can fire off thousands of model calls per day. At the going rate for the most powerful models, that adds up fast. When a specialist model costs 95% less, those same workflows become affordable to run all day, every day.
The practical question is no longer just which AI model is most capable. For routine tasks — code reviews, document summaries, security scans, customer queries — a cheaper specialist is usually good enough. Save the expensive flagship model for genuinely hard, one-off problems. The gap between the best available and the best value just got much wider.