Microsoft unveils new models to cut agent operating costs

Microsoft has introduced seven specialized models for code, security, images, speech, and workflow routing designed to lower operating expenses for always-on AI agents.

2 min read

Microsoft unveils new models to cut agent operating costs

Microsoft's AI division has unveiled seven specialized models aimed at reducing the operating costs of always-on AI agents. The release includes models tailored for software development, IT security, image editing, speech recognition, and workflow decisions, positioning them as lower-cost alternatives to systems from OpenAI and Anthropic.

Key facts

Context window
1,000,000 tokens
MAI Code 2 input price
$0.50 per million input tokens
MAI Code 2 output price
$2.50 per million output tokens
MAI Cyber 2 Flash input price
$0.10 per million input tokens
MAI Cyber 2 Flash output price
$0.50 per million output tokens
Coding Model Performance and Cost Comparison
Promotional input price
MAI Code 2: $0.50 per million tokens
MAI Code 2 Flash: $0.05 per million tokens
Promotional output price
MAI Code 2: $2.50 per million tokens
MAI Code 2 Flash: $0.25 per million tokens
Standard input price
MAI Code 2: $1 per million tokens
MAI Code 2 Flash: $0.10 per million tokens
Standard output price
MAI Code 2: $5 per million tokens
MAI Code 2 Flash: $0.50 per million tokens
Bar charts showing performance metrics and cost comparisons between MAI Code 2, GPT-6.1 Sol, and Claude Opus 5.5.
Bar charts comparing MAI Code 2, GPT-6.1 Sol, and Claude Opus 5.5 across performance benchmarks and cost per task., Source ↗

New Coding and Work Models

The coding lineup is led by MAI Code 2, which Microsoft reports reaches 98 percent of the performance of GPT-6.1 Sol across four benchmarks at 40 percent of the cost. A smaller variant, MAI Code 2 Flash, is designed for routine fixes and explanations. Additionally, MAI Work Flash integrates into Microsoft 365 and Fabric IQ, handling tasks across Excel, PowerPoint, and Word while consuming fewer input tokens.

Security and Routing Updates

For IT security, MAI Cyber 2 Flash resolves tasks in the CyberGym-E2E vulnerability benchmark while operating at a fraction of the per-task cost of comparable models. Its context window spans one million tokens, marking a fourfold increase over its predecessor.

Bar charts showing MAI Cyber 2 Flash outperforming GPT-6 Astra at a fraction of the cost.
Bar charts comparing MAI Cyber 2 Flash and GPT-6 Astra on task success, output tokens, and cost per task on CyberGym-E2E., Source ↗

For workflow routing and classification, the Decision 1.1 model is built on Google's mixture-of-experts model Gemma 4 26B A4B — a machine learning architecture that routes inputs to specialized sub-networks rather than activating the entire network at once. It achieves faster execution and lower costs compared to prior versions.

Scatter plot comparing Microsoft-Decision-1.1 performance and speed against competing models.
Scatter plot showing the Artificial Analysis Decision Index score versus time per task for Microsoft-Decision-1.1 and other models., Source ↗

Media and Speech Capabilities

In image editing, MAI Image 2.7 features object manipulation capabilities and transparent background generation while leading specific evaluation leaderboards. Meanwhile, MAI Realtime 1 serves as Microsoft's first speech-to-speech model, featuring compatibility with the OpenAI Realtime API to support voice applications and customer service workflows.

Enterprise customers will also be able to fine-tune selected model weights using their own data within the Azure cloud environment, keeping customizations secure.

Source: Original announcement

Worth a read?

Report

Comments