Subscribe
AI Models

Microsoft unveils seven in-house AI models built to undercut OpenAI and Anthropic on cost

Wednesday, November 18, 2026 · 6 min read

At Ignite 2026, Microsoft launched seven task-specific models — including a coder and a cyber defender — priced to do each job for a fraction of what rivals charge.

Bullish$MSFT

📊 Scorecard: our Bullish call on $MSFT is scored from $506.24, the last close before it ran. Its first result comes after the next close.

Microsoft unveils seven in-house AI models built to undercut OpenAI and Anthropic on cost

The News

Microsoft unveiled a family of seven task-specialised models at Ignite 2026 in San Francisco on Tuesday, November 17, its biggest push yet to compete with OpenAI and Anthropic on cost per job rather than raw capability. The headliners are MAI Code 2 for enterprise coding, available in GitHub Copilot with a Flash variant in limited preview in Microsoft Foundry, and MAI Cyber 2 Flash for defensive cybersecurity, initially limited to a small group of customers.

The rest of the family covers image generation and editing (MAI Image 2.7), live speech (MAI Realtime 1, Microsoft's first live speech-to-speech model, from $0.01 per audio minute), transcription (MAI Transcribe 2), and Decision 1.1, a router that sends each request to a right-sized model. GPT-6.2 and Meta's latest models are also coming to Foundry.

Microsoft says Code 2 averages 98% of GPT-6.1 Sol's score across four coding benchmarks at about 40% of the price — but that price is a 50% promotional rate, so at full price Code 2 costs roughly 80% of GPT-6.1 Sol, and Cyber 2 Flash beats GPT-6 Astra on CyberGymE2E at under 2% of the cost per task — though Microsoft excluded 92 of the benchmark's 905 tasks because they overlapped its training data. Those are Microsoft's own claims, and no independent scores exist yet. Microsoft AI chief Mustafa Suleyman told Reuters the domain-specific strategy keeps costs low while matching rivals.

Microsoft is not trying to build the smartest model. It is trying to build the cheapest one that is good enough.

Why It Matters

The AI market is shifting from 'which model is smartest' to 'which model does this job cheapest'. Most business work does not need a frontier model — it needs a reliable one that costs pennies. A specialised model that scores 98% as well for 40% of the price at a 50% promotional rate — about 80% at full price — wins that argument almost every time, because the missing 2% rarely matters for routine code. The same logic applies to images, speech and transcription: each of those jobs is routine enough that a small model can handle it, and the savings add up across every request a company makes.

The router, Decision 1.1, is the quietest and maybe the most important piece. It sends each request to a right-sized model, so a simple question never pays frontier prices. That is how Microsoft turns seven models into a system: cheap models handle the bulk of the work, and the expensive ones only get called when they are needed.

The risk is that Microsoft is grading its own homework. Its benchmark claims come from its own labs, and the history of AI launches is full of vendor numbers that outside testers later trimmed. Until Artificial Analysis or a similar independent tester weighs in, the 98% and under-2% figures should be read as a direction, not a fact.

Analyst's Take

Our call is Bullish. If even half of Microsoft's claims hold up, the economics of AI change: a coder at 98% of the rival's score for 40% of the price — but that price is a 50% promotional rate, so at full price Code 2 costs roughly 80% of GPT-6.1 Sol, and a cyber defender at under 2% of the cost per task, would let companies run far more AI for the same budget. Microsoft is selling the work, not the wonder, and that is where the money is.

The honest caveat is that none of this has been independently checked yet. Microsoft's own benchmarks are a promise, not proof. And OpenAI launched GPT-6.2 the same day, so Microsoft's comparison against GPT-6.1 Sol is already one generation behind. The claims are plausible — a small model built for one job should be cheap — but they are still claims.

For readers, the thing to watch is Foundry, Microsoft's model marketplace. If customers start picking MAI models for routine jobs and saving the expensive frontier models for hard ones, Microsoft wins twice: it sells its own models and still hosts its rivals'. That is a strong position either way. The test will come in the next few months, when independent testers publish their own scores and customers report what the models actually cost in production. Reader poll: Bullish or bearish on Microsoft's AI bet? Vote: Bullish, Bearish, Mixed.

Are you bullish or bearish on this story?

By The Numbers

7Task-specialised models unveiled
98%Code 2's average score vs GPT-6.1 Sol, Microsoft's claim
40%Code 2's price vs GPT-6.1 Sol at a 50% promotional rate (~80% at full price), Microsoft's claim
<2%Cyber 2 Flash cost per task vs GPT-6 Astra, Microsoft's claim
$0.01MAI Realtime 1 price per audio minute
$506.24MSFT close, Nov 17, -2.27%

FAQ

What is a foundation model?

A foundation model is a large, general-purpose AI model trained on huge amounts of data, which can then be adapted to many different jobs. The new Microsoft models are the opposite idea: smaller models trained for one specific job, like coding or transcribing speech.

What is GitHub Copilot?

GitHub Copilot is an AI coding assistant that suggests code as a programmer types. It is owned by Microsoft, and it is one of the biggest ways Microsoft sells its AI models to developers.

What does 'cost per job' mean?

Cost per job is what a customer pays for the AI to complete one piece of work, like writing a function or checking a network for intruders. Microsoft's bet is that a small model built for one job can do it nearly as well as a giant model, for much less money.

👁

Source: 943 Jack FM

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $MSFT

All $MSFT coverage →

More from AI Agent Update