Microsoft bets on efficiency: seven new MAI models take on frontier costs
On November 17, 2026, in an unsigned post on its official website titled “Seven new world-class models at scalable price-performance”, Microsoft AI announced a lineup of seven proprietary models in the MAI family. The list includes MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Realtime 1, Decision 1.1 and Dragon Copilot Transcribe, a product built on the MAI Transcribe 2 engine. During the Ignite presentation, according to Thurrott.com, Microsoft AI CEO Mustafa Suleyman summed up the approach: “We don’t just hill-climb on frontier quality alone. We co-optimize quality and price together.” The launch expands the MAI family, which Microsoft develops in-house alongside its partners' models, and uses OpenAI's and Anthropic's products as the yardstick for pricing and performance.
All the performance figures come from the company itself and have not yet been independently verified. In coding, MAI Code 2 averages 98% of GPT-6.1 Sol's quality at roughly 40% of the price, and 96% of Claude Opus 5.5's performance at 5% of the price, measured across four industry benchmarks. MAI Code 2 Flash performs comparably to GPT-6 Luna at 65% of the price. According to the model card, MAI Code 2 Flash has about 12 billion active parameters out of roughly 398 billion in total, and both Code 2 models have a 1,050,000-token context window. The documentation also notes that the coding models are not optimized for complex tasks such as fundamental scientific research and long-horizon agentic work. In cyber defense, Microsoft says MAI Cyber 2 Flash beats GPT-6 Astra on the CyberGym-E2E benchmark at less than 2% of the cost per task, while Thurrott.com reports the two as on par.
On the multimedia and operations side, MAI Realtime 1 is the company's first real-time speech-to-speech model, priced from $0.01 per minute of audio. Alongside it come MAI Image 2.7, ranked first on the Artificial Analysis image-editing leaderboard at one fifth of the price of GPT Image 2.6 Sunburst (max), and Decision 1.1 for request routing. Microsoft says Maia 200, its second-generation AI accelerator, is now in production and running its largest models, with 1.7 times more tokens per watt. As for distribution, MAI Code 2 is built into GitHub Copilot, and MAI Code 2 Flash is also available as a limited-access preview on Microsoft Foundry. The launch post does not specify release regions, though, so availability in Europe is still undefined. In a testimonial published by Microsoft, Abhay Dabholkar, AVP of Data Science at AT&T, says the customized version of Code 2 Flash used by AT&T outperforms GPT-6.1 Sol in standard 5G network analysis at one fiftieth of the cost.
The pursuit of raw performance seems to be giving way to a colder, more sustainable calculation about compute. Cutting token costs without giving up too many benchmark points is a real need for anyone building software at scale. Still, how solid these numbers are will only become clear through independent testing, beyond the metrics the company supplies itself. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official microsoft.ai post on Ignite 2026 and the MAI Code 2 model page, and took the names, figures and quotes from them. I compared the data with Paul Thurrott's independent article of 17/11/2026, which confirms the seven names and Suleyman's quote, and noted the differences. I found the Microsoft blog post and the PDF model card but did not reread them in full. I left out topics already covered (GPT-6.2, Gemini 4 Hydrogen), stories with no primary source (AMD–World Labs, Isomorphic Labs) and stories outside the 7-day window (Tencent Hy4, November 4).
- Incertezze
- The performance and price comparisons are Microsoft's own claims and have not been independently verified. Almost all prices are relative percentages: only Realtime 1 and Image 2.7 Flash have dollar prices. The sources disagree on a few points. On Cyber 2 Flash, Microsoft says it “beats” GPT-6 Astra, while Thurrott says it “matches” it. On Code 2 Flash versus Haiku 5.5, the blog says performance is “comparable” at 20% of the price, while the model page says “12% better” at one fifth of the cost. Thurrott says all the models run on Maia 200, while Microsoft only mentions “our largest models”. The “seven models” include Dragon Copilot Transcribe, a product built on MAI Transcribe 2, and leave out MAI Image 2.7 Flash, even though the text mentions it. No regions are given, so availability in Italy and the EU is unclear. For Decision 1.1, the post does not explicitly say “general availability”.
- Perché pubblicarla
- This is Microsoft's broadest release of proprietary models, announced at Ignite a few days ago. One of the largest AI distributors is moving to its own low-cost models and competing head-on with OpenAI and Anthropic. The release affects tools that Italian developers and companies use, such as GitHub Copilot and Foundry. The site had not covered it yet.
Fonti / Sources
- Microsoft AI – Seven new world-class models at scalable price-performance (annuncio ufficiale, 17/11/2026)
- Microsoft AI – pagina modello MAI Code 2 and Flash
- Microsoft AI – Model card MAI Code 2 / MAI Code 2 Flash (PDF, aggiornata 17/11/2026)
- Official Microsoft Blog – Ignite 2026: Building the frontier ecosystem