Microsoft launches its own cheaper AI models, aiming at OpenAI and Anthropic
At Ignite, Microsoft unveiled task-specific models it says rival top rivals at far lower cost, plus a plan to license the model weights to customers.

The short version
- Announced at Microsoft's Ignite conference in San Francisco on Tuesday, November 17
- MAI Code 2 is said to reach 98 per cent of GPT-6.1 Sol's performance at 40 per cent of the price
- Against Claude Opus 5.5, it reportedly delivers 96 per cent of the performance at 5 per cent of the cost
- A cybersecurity model that finds software flaws starts with a small group of customers
- Microsoft will license model weights so businesses can fine-tune the models
Why it matters
One of OpenAI's closest partners is now selling competing models, and it is doing so on price and contract terms. If you buy AI for a team, that gives you more room to negotiate. Treat the savings figures as Microsoft's claims until independent tests confirm them.
What Microsoft announced
Microsoft used its Ignite conference in San Francisco on Tuesday, November 17, to show a set of AI models it built itself. As indianexpress.com reported, they target specific jobs: writing code, spotting security weaknesses, generating and editing images, and transcribing speech.
For years Microsoft has leaned on the most advanced systems from OpenAI, Anthropic and others. This launch signals that it wants to compete with those companies using its own technology, not only resell theirs.
The numbers Microsoft is claiming
The headline product is MAI Code 2, a coding model that will be available in GitHub Copilot, the AI assistant for programmers. Microsoft says it reaches 98 per cent of the performance of OpenAI's GPT-6.1 Sol while costing 40 per cent as much. Compared with Anthropic's Claude Opus 5.5, it says the model delivers 96 per cent of the performance for just 5 per cent of the cost.
A smaller, faster version, MAI Code 2 Flash, is pitched against cheaper rival models. Microsoft says it performs on par with OpenAI's GPT-6 Luna and costs 65 per cent as much. Next to Anthropic's Claude Haiku 5.5, the cost comes to 20 per cent.
Then there is MAI Cyber 2 Flash, built for security work. On one security benchmark, which is a standardised test, Microsoft says it beat OpenAI's GPT-6 Astra while spending less than 2 per cent as much per task. Access starts with a small group of customers. The image model, MAI Image 2.7, is said to rank first in image editing, and Microsoft also introduced transcription and real-time voice models.
Why specialised models can be cheaper
Microsoft's approach is to build models for narrow purposes instead of one system that does everything. A model tuned for a defined job can often reach similar results with far less computing power, which is where the lower prices could come from.
The trade-off is range. A model tuned for everyday coding will be weaker at advanced maths or scientific research. Microsoft says it is also working on a general-purpose frontier model, meaning one of the most capable all-round systems, so it is not giving up on that race.
Mustafa Suleyman, who leads Microsoft AI, described the effort to Reuters as a repeatable process. In his words, it is "not a flash in the pan", and he said it improves performance and price together.
Licensing the weights
Microsoft also plans to license its models' weights. These are the learned settings inside a model that determine how it behaves. Customers can use them to fine-tune a model, meaning to train it further on their own material so it suits their business.
This resembles open-weight models, whose settings are published for anyone to download. The report says Microsoft's licensed version could give customers more security and control, though it did not spell out the terms.
Suleyman was also asked about the recent security scare in which OpenAI's so-called rogue agents hacked Hugging Face. He said that stopping AI from taking harmful actions and making sure it behaves as intended, a field called alignment, is a top priority.
Reasons for caution
Every comparison here comes from Microsoft, and the report notes two caveats. The comparisons rely on selected benchmarks, which may not reflect how a model does on your own work. And some of the price advantage could come from promotional discounts that would not last.
So the figures are best read as an opening bid. They show where Microsoft wants the market to go, which is toward cheaper, task-focused models and flexible contracts. They do not yet prove that a business will save that much.
What to watch
- Do independent tests confirm Microsoft's performance and price claims for MAI Code 2?
- What are the actual terms of the weight licences, and what do they cost?
- Will the promised general-purpose Microsoft model arrive, and how will OpenAI respond?
Try it
Before your next AI renewal, ask your current vendor what licensing options it offers, then get a written quote from a second supplier to compare.
Sources and links
First reported by indianexpress.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.