Skip to content
Apixo
Blog
news· 4 min read· via The Decoder

Microsoft Unveils Seven In-House AI Models to Cut Costs on Continuous Agent Workloads

Microsoft AI has launched seven specialized models for coding, routing, security, and speech, targeting lower token costs for enterprise agents compared to OpenAI and Anthropic.

Microsoft Unveils Seven In-House AI Models to Cut Costs on Continuous Agent Workloads

Microsoft AI has introduced a suite of seven new specialized models covering coding, cybersecurity, decision routing, vision, and speech. Announced in a single release at the company's Ignite conference, the lineup is explicitly engineered to tackle the ballooning expense of running autonomous, always-on software agents. Rather than positioning these systems for foundational research or intricate long-horizon logic, Microsoft is aiming squarely at everyday enterprise workloads where continuous token consumption quickly strains corporate budgets.

The initiative signals a calculated effort by Microsoft to provide cost-effective first-party options alongside its offerings from OpenAI, giving organizations tools to handle repetitive and operational agent tasks for a fraction of the price.

Code and security at a fraction of the cost

Leading the software engineering side is MAI Code 2. Microsoft reports that the model matches 98 percent of GPT-6.1 Sol's performance across four benchmarks—DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1—at 40 percent of the cost. When compared to Claude Opus 5.5, Microsoft states it hits 96 percent of the performance at 5 percent of the price. The tests were run without sub-agents on Microsoft's MSBench platform using the GitHub Copilot CLI. In blind evaluations of frontend tasks like SVG graphics and user interfaces, human evaluators picked MAI Code 2 over GPT-6 Sol 64 percent of the time.

Through a promotional rate running until May 17, 2027, MAI Code 2 is priced in GitHub Copilot at $0.50 per million input tokens and $2.50 per million output tokens, after which rates adjust to $1 and $5. The lighter MAI Code 2 Flash targets routine edits and code explanations at $0.05 per million input tokens and $0.25 per million output tokens during the promotion. Microsoft claims it equals GPT-6 Luna and Claude Haiku 5.5 at 65 percent and 20 percent of their respective costs.

For enterprise defense, Microsoft launched MAI Cyber 2 Flash, which solves 92.3 percent of problems on the CyberGym-E2E benchmark versus 91.5 percent for GPT-6 Astra, while requiring half the reasoning budget. Priced at $0.10 per million input tokens and $0.50 per million output tokens, Microsoft reports that per-task costs fall from $34.38 on GPT-6 Astra to $0.45 on the MAI model. The security model features a one-million-token context window and already manages roughly 40 percent of scans on Microsoft's MDASH platform.

Specialized routing, vision, and voice

Microsoft also introduced Microsoft-Decision-1.1, an operational routing engine built on Google's Gemma 4 26B A4B mixture-of-experts architecture. Designed to handle workflow control, prioritization, and confidence scoring, it directs queries inside Copilot 2.2 times faster and 4.4 times cheaper than standard large language models. The engine accepts image inputs and costs $0.042 per million input tokens ($0.05 in the EU), while output tokens are free.

For retrieval, Microsoft-Retriever-1 enters preview in Foundry IQ to iteratively plan and evaluate queries across internal documents. In computer vision, MAI Image 2.7 topped the Artificial Analysis image-editing leaderboard at 1,215 Elo, outscoring GPT Image 2.6 Sunburst while costing $35 per 1,000 images, or $10 for the Flash tier.

In audio, MAI Realtime 1 marks Microsoft's first speech-to-speech model. Compatible with OpenAI's Realtime API, the system runs on Microsoft's Maia 200 hardware, charging one cent per audio minute alongside optional per-token reasoning fees.

What it means for developers

For engineering teams architecting production agents, Microsoft's batch release highlights a major industry shift toward multi-model systems. Relying on a single top-tier frontier model for every subtask—such as classification, document retrieval, and basic code updates—is becoming financially unsustainable as agents run around the clock.

By releasing dedicated routing models like Decision 1.1 and API-compatible voice tools, Microsoft is making it easier to offload high-frequency steps to cheaper, specialized infrastructure while preserving high-end reasoning models for complex blockers. Furthermore, the compatibility of MAI Realtime 1 with OpenAI's endpoints shows that switching costs between model providers are falling.

Because production stacks increasingly mix and match vendors based on performance benchmarks and price per token, developers can try top AI models cheaply through one API at https://apixoai.online to evaluate alternative architectures. As enterprise platforms like GitHub Copilot, Vercel, and OpenRouter integrate these lower-cost alternatives, developers have more flexibility than ever to optimize agent operating costs without completely rewriting their underlying workflows.


Source: Microsoft is replacing outside AI models in its own products with cheaper in-house alternatives — The Decoder. Written by the Apixo team from that report.

#ai-news#microsoft#ai-models#github-copilot#openai#benchmarks
Try it with your own tools

One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.

Get your API key

Keep reading