Skip to content
ISSUE 001·LIVE·13:04 IL
← Journal/2026-11-19·9 min·microsoft

Microsoft's seven MAI models: the price-performance play for 24/7 coding agents

At Ignite 2026, Microsoft AI released seven models led by MAI Code 2, which it claims reaches 98% of GPT-6.1 Sol's coding quality at 40% of the price, plus a cyber model at under 2% of a frontier competitor's cost per task, a routing model, and Licensed Weights you run in your own Azure tenant. The strategy is clear: not benchmark trophies, but agents cheap enough to run around the clock.

By Harel Asaf·AI Builder·Tel Aviv

Microsoft AI shipped seven new models at Ignite this week, and the announcement is worth reading as a strategy document, not a benchmark dump. The pitch: as token spend skyrockets, enterprises want agents running real engineering work 24/7, and that needs models at a price point where continuous operation is affordable. So MAI Code 2 claims 98% of GPT-6.1 Sol's coding quality at about 40% of the price, MAI Cyber 2 Flash claims to beat GPT-6 Astra at under 2% of the cost per task, and MAI Code 2 is already in GitHub Copilot today. Every comparison below is Microsoft's own claim from its announcement posts; none of it is independently verified yet, and that caveat matters more than usual because every number here is a price-performance ratio where Microsoft picks both the benchmark set and the reference price.

The lineup at a glance

Seven models: MAI Cyber 2 Flash (security agents), MAI Code 2 and MAI Code 2 Flash (everyday enterprise coding), MAI Image 2.7 and 2.7 Flash (image editing and generation), MAI Realtime 1 (speech-to-speech, from $0.01 per audio minute), MAI Transcribe 2 (powers Dragon Copilot Transcribe at 10 million patient encounters a month), and Decision 1.1 (model routing). Alongside them: Maia 200, Microsoft's second-generation accelerator, now in production with a claimed 1.7x more tokens per watt. This is a full-stack announcement: silicon, models, routing, and distribution through Copilot and Foundry.

MAI Code 2: honest about what it is for

The coding models are the interesting part for builders, partly because Microsoft is unusually explicit about scope. MAI Code 2 averages its claims over four benchmarks targeting everyday enterprise coding seen in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. Against Claude Opus 5.5 it claims 96% of the performance at just 5% of the price. In a blind side-by-side evaluation of frontend work, developers reportedly preferred MAI Code 2 over GPT-6 Sol 64% of the time, and Microsoft says the model is tuned to produce good dashboards and visualizations.

Then the honest part, quoted almost directly: these models are explicitly not yet tuned for harder problems like fundamental science research and very long horizon agentic work. That is a rare admission in a launch post, and it tells you exactly where to point MAI Code 2 (high-volume routine tasks, targeted edits, explanations, iterations) and where not to (your autonomous overnight refactoring agent). Code 2 Flash goes further down-market: comparable to GPT-6 Luna at 65% of the price and Haiku 5.5 at 20%, aimed at the highest-volume work.

Cyber, routing, and the rest

MAI Cyber 2 Flash continues Microsoft's bet that security monitoring is the killer always-on workload: defense and patching agents watching live production systems continuously. Microsoft says MAI Cyber already accounts for almost 40% of scans in MDASH (its defense platform), 90% of scans that started with MAI kept using it, and the new model cuts output-token price 86% versus Cyber 1.1 Flash. Decision 1.1 is a routing model that sits in front of your application and sends requests to expensive models only when needed; Microsoft claims the top spot on the Artificial Analysis Decision Index, 4,000 customers, and 50x faster and 79x cheaper than GPT-6.1 Sol (an odd comparison for a router, but the customer count is the real signal).

Licensed Weights: the actual strategic move

Buried in the lineup is the most enterprise-shaped announcement: MAI Licensed Weights. Take a frontier MAI model, customize it on your proprietary data, run it inside your own Azure environment, control access, keep your adaptations. Microsoft's framing is a direct shot at open weights ("can you really trust where they came from?") and at closed API rivals ("no one wants to pour their IQ into someone else's model"). The proof point is AT&T: a custom MAI model for 5G network operations analytics that, Microsoft says, outperformed GPT-6.1 Sol at 1/50th of the cost. Delivery runs through Microsoft's Frontier Company program. If this works, it is a new shape of model distribution: not API rental, not open download, but licensed custody.

What I could not verify

Every comparative number is Microsoft's, from benchmarks it selected, at reference prices it computed. There are no third-party evaluations of any of these models yet, the per-model posts do not publish full eval methodology in what I reviewed, and claims like "64% developer preference" and "96% of Opus 5.5 at 5% of the price" need independent reproduction before they justify migration. The scope caveat on long-horizon work is Microsoft's own. I also could not verify current Copilot model-picker availability from the posts alone.

Frequently asked questions

What models did Microsoft release at Ignite 2026?

Seven: MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Image 2.7 Flash, MAI Realtime 1, MAI Transcribe 2 (behind Dragon Copilot Transcribe), plus the Decision 1.1 routing model and the Maia 200 accelerator going into production.

How good is MAI Code 2 compared to GPT-6.1 Sol and Claude Opus 5.5?

By Microsoft's own measurements across four everyday-coding benchmarks, 98% of GPT-6.1 Sol's quality at about 40% of the price, and 96% of Opus 5.5's quality at about 5% of the price. These are vendor claims awaiting third-party verification.

Where can I use MAI Code 2 today?

In GitHub Copilot, where Microsoft says it is available now. MAI Code 2 Flash is in Copilot plus a limited-access preview in Microsoft Foundry. The image models are in Foundry, and Realtime 1 is in MAI Labs.

What are MAI Licensed Weights?

A licensing program where you take a frontier MAI model, optionally fine-tune it on your data, and run it inside your own Azure environment with your access controls. Your data and adaptations stay yours; Microsoft delivers through its Frontier Company program. It is positioned against both open weights and pure API rental.

Is MAI Code 2 meant for long-horizon agentic work?

No, and Microsoft says so explicitly: the coding models are tuned for everyday enterprise tasks and are not yet aimed at fundamental research or very long horizon agentic work. Use them for high-volume routine coding, not autonomous multi-hour runs.

How much does MAI Realtime 1 cost?

It starts at $0.01 per audio minute, and Microsoft claims it outperforms GPT-Realtime 2.1 on external and Dynamics 365 benchmarks. It is available to try in MAI Labs.

Should I switch my coding agents to MAI models?

If you already live in Copilot or Azure, MAI Code 2 is worth benchmarking on your routine tasks now since it is already in the Copilot model picker. For anything outside Microsoft's ecosystem, wait for independent evals; the entire value proposition is a price-performance ratio that only third parties can confirm.

Sources

Related reading

  • GPT-6.2 Astra and Sol: OpenAI's answer to the price-performance war

    OpenAI released GPT-6.2 Astra and Sol on November 17, 2026: a frontier model for the hardest professional work at $8 per million input tokens, and an everyday model holding GPT-6.1 Sol's prices. The launch leans on cost-per-task claims across legal, science, coding and 3D benchmarks, ships both models in ChatGPT Work, Codex and the API, and includes a rare disclosure: Astra is OpenAI's first external model to hit the High threshold for AI self-improvement.

  • Gemini 4 Neon: Google's fast reasoning model for production coding agents

    Google launched Gemini 4 Neon on November 18, 2026: a fast reasoning model positioned between Carbon and Hydrogen, priced at $3 per million input tokens with a 1M context window. Google claims #1 on the Artificial Analysis Coding Agent Index, state-of-the-art SWE-sweep results inside Antigravity, and 32% lower cost per task than Carbon. Here is what shipped, what the benchmarks say, and what to watch.

  • Qwen4-Max: Alibaba's 2.4T flagship API model claims 62% lower cost per task than its predecessor

    Alibaba's Qwen team released Qwen4-Max, the first model in the Qwen4 family: a 2.4-trillion-parameter sparse MoE with 71B active parameters, a 1M-token context window, thinking modes with tunable effort, and both OpenAI-compatible and Anthropic-compatible endpoints. Vendor-reported numbers show up to 49% fewer tokens on agentic coding tasks and a 62% drop in average cost per task versus Qwen3.8-Max. Input starts at 1.5 dollars per million tokens.

The build log

Follow the work.

New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.

Follow via RSS