Microsoft used the opening keynote of its Ignite conference in San Francisco on Tuesday to put six new in-house MAI models on the table, and the pitch was less about a single breakthrough than about a systematic attack on price. Mustafa Suleyman, chief executive of Microsoft AI, framed the family — spanning coding, cybersecurity, image generation, transcription and real-time voice — as proof that domain-tuned models can match frontier systems on selected benchmarks at a fraction of the cost.
The centerpiece is MAI Code 2, a general-purpose coding model built for the work developers actually do: deep debugging, multi-file changes, repository-wide refactoring and front-end work. Microsoft says that averaged across four coding benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — MAI Code 2 delivers 98% of the performance of OpenAI's GPT-6.1 Sol at 40% of the price, and 96% of Anthropic's Claude Opus 5.5 at just 5% of the cost. In blind tests of front-end apps, the company says reviewers preferred its output over GPT-6 Sol roughly two-thirds of the time. Both figures are Microsoft's own claims, based on selected benchmarks, and the cost comparisons lean on the models' 50% promotional pricing — meaning the gap narrows once the promotion ends.
A smaller sibling, MAI Code 2 Flash, targets high-volume everyday tasks like code explanation, quick edits and routine bug fixes. Microsoft positions it as matching GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20% of the cost. Both coding models are live now in the model picker in GitHub Copilot across paid plans, with Code 2 Flash also available on Copilot Free and Student tiers and in a limited-access Microsoft Foundry preview. OpenRouter, Vercel, Hermes, OpenCode and Kilo integrations are promised next. Microsoft has also tuned the Flash variant into MAI Work Flash to serve Word, Excel, PowerPoint, Teams and Fabric IQ.
The boldest claim of the day came from the security model. MAI Cyber 2 Flash, which Microsoft calls its most capable security model with 12 billion active parameters, is said to match GPT-6 Astra's quality on the CyberGymE2E benchmark at less than 2% of the cost per task. The company plans to offer it to a limited group of customers, citing a surge in software vulnerabilities that AI tools are surfacing faster than humans can patch. Rounding out the lineup, Microsoft says MAI Image 2.7 ranks first in image editing by its own evaluation, and Microsoft-Decision-1.1 handles the small routing and verification choices agents make thousands of times per run.
Suleyman told Reuters the strategy is deliberate specialization rather than a race for a single general model: "This isn't just one model. It's not a flash in the pan. This is a system, a hill-climbing machine that we have built, which co-optimizes performance and price." He acknowledged the trade-off — a model tuned for everyday coding will not reach the frontier on fundamental science or very long-running agent tasks — and confirmed a general-purpose frontier model is in progress without a timeline. Enterprise customers can license the model weights and deploy them on Azure with their own data.
The timing was no accident. OpenAI released GPT-6.2 Astra and Sol on the same day, cutting the Astra API price by 20%, so the two companies ended up competing openly on price within hours of each other. Ignite also carried broader news — an AI Transparency Initiative with NIST and more than 40 organizations, a Copilotforce partnership bringing Salesforce data into Copilot and Teams, and Cobalt 200 virtual machines reaching general availability — but the model family is the part with a direct line to Microsoft's billings.
The underlying bet is commercial rather than academic: every task MAI Code 2 absorbs inside Copilot is one that no longer pays a rival's API bill. Whether the claimed ratios survive independent benchmarking matters less than the direction of travel — for enterprise buyers, "good enough at a twentieth of the cost" is becoming a purchasing category, and both OpenAI and Microsoft spent Tuesday leaning into it.
Comments (0)
Log in to join the discussion
Log InNo comments yet