Microsoft strengthens its in-house models to reduce AI costs
Microsoft introduces a range of in-house models for code, security, image, voice, and routing. The goal is clear: reduce the cost of AI agents while maintaining enterprise use within Azure.
Microsoft unveiled seven specialized AI models at Ignite, designed for code, security, image, voice, and automated decision-making. The challenge is both financial and technical: the company aims to replace some OpenAI and Anthropic models with cheaper alternatives in continuously running products. This strategy matters for businesses, as AI agents quickly consume token volumes that significantly impact costs.
A homegrown range designed for nonstop use cases
The most visible novelty is not a single model, but a complete portfolio. Microsoft wants to cover several components at once: code generation, security support, image editing, voice processing, task routing, and document search. The models were unveiled together at the Ignite conference, following an initial release at Build in June, with a clear focus on cost-performance ratio.
For code, MAI Code 2 targets common enterprise tasks. Microsoft states it achieves on average 98% of GPT-6.1 Sol’s performance across four internal tests, at 40% of the cost, and 96% compared to Claude Opus 5.5 at 5% of the cost. However, the company notes this model is not yet optimized for in-depth research or long-duration agent workflows. In short, it codes well but does not yet claim to rewrite the entire software plumbing.
MAI Code 2 Flash is aimed at code explanations, targeted corrections, and routine adjustments. Microsoft positions it at 65% of GPT-6 Luna’s cost and 20% of Claude Haiku 5.5’s, according to its own measurements. In GitHub Copilot, the promotional price announced is $0.50 per million input tokens and $2.50 per million output tokens, valid until May 17, 2027.
Secutiry, image, and voice: Microsoft seeks substitution where costs are high
On cybersecurity, MAI Cyber 2 Flash is presented as more efficient and significantly less costly than GPT-6 Astra on CyberGym-E2E. Microsoft reports 92.3% of tasks solved, versus 91.5% for GPT-6 Astra, with half the reasoning budget and a cost per task under 2% of the competitor’s. On WebExploitBench, the model scores 40.4%, ahead of GPT-5.6 Cyber at 38.4%. The context is sensitive: Microsoft also states it removed from its evaluation tasks related to open-source projects and vulnerability reports present in its training data. [to be verified]
MAI Image 2.7 tops an image editing ranking by Artificial Analysis, with 1,215 Elo points and 5,478 votes, four points ahead of GPT Image 2.6 Sunburst. The model can modify color, material, or position of an object without affecting the rest of the image, and generate transparent backgrounds. Its announced price is around $35 for 1,000 images, while the Flash version drops to about $10.
With MAI Realtime 1, Microsoft enters the real-time speech domain. This first speech-to-speech model from the brand handles interruptions, waits for the speaker to finish, and recognizes requests despite background noise. It is compatible with the OpenAI Realtime API, facilitating migration of existing applications. Tests are already underway at Agoda and Reservations.ai, while Dynamics 365 Contact Center and Dragon Copilot are targeted.
Routing becomes a profession in its own right
Microsoft is not only betting on models that produce text or images. With Decision 1.1, the company pushes a routing model capable of classifying, prioritizing, and directing workflows. The system chooses among developer-approved actions, returns a confidence score, and can abstain if information is lacking. It is a discreet but strategic piece: in many products, deciding what to do matters as much as responding quickly.
Decision 1.1 is based on Google’s Gemma 4 26B A4B, following a predecessor built on Qwen3.5-9B. Microsoft says it trained it with supervision and reinforcement learning on synthetic and licensed datasets. On Artificial Analysis’s Decision Index v2.1, it scores 68 points, six more than the previous version. In Copilot, it notably helps decide whether a request should go to Copilot or Cowork.
The model also accepts images, including screenshots. Microsoft claims it is as accurate as a large language model but 2.2 times faster and 4.4 times cheaper. The announced price is $0.042 per million input tokens, or $0.05 in the European data zone, with output tokens free.
Licensed weights to keep data in-house
Another important announcement: Microsoft launches "Licensed Weights," a system allowing companies to fine-tune certain MAI models with their own data within their Azure environment. The idea is to keep data and adjustments on the client side while avoiding the gray areas of open weights, whose origin and behavior are harder for companies to verify.
Microsoft cites AT&T as an example, where a custom model for 5G network operations analysis reportedly outperformed GPT-6.1 Sol on standard tasks at about one-fiftieth of the cost. The service has no launch date or public price yet; interested companies must contact their Microsoft sales team. The offer is expected "shortly," via the Frontier Company program.
This offensive fits into a more flexible relationship with OpenAI. In April, the two companies reorganized their partnership: OpenAI can now sell its products through any cloud provider, although Azure remains the main partner. Microsoft retains a non-exclusive license on OpenAI’s models and products until 2032. In this context, multiplying in-house models is not just a technical move. It is also a way to regain control over margin and pace.
Key points
- Seven MAI models announced at Ignite, November 2026.
- MAI Code 2 targets 98% of GPT-6.1 Sol at 40% of the cost.
- MAI Cyber 2 Flash reaches 92.3% on CyberGym-E2E.
- MAI Image 2.7 claims 1,215 Elo points.
- Decision 1.1 scores 68 points on Decision Index v2.1.
- Microsoft maintains a commercial link with OpenAI until 2032.
By the numbers
- 7 models — announced by Microsoft at Ignite, November 2026.
- 1 million tokens — context window of MAI Cyber 2 Flash.
- 10 million — monthly conversations transcribed by Dragon Copilot.
- $0.05 — price per input token of Decision 1.1 in the EU.
- May 17, 2027 — end of MAI Code 2 promotional pricing.