Microsoft announced MAI Code 2 and six related models at its Ignite conference on November 17, 2026, positioning the family as a lower-cost alternative for enterprises running agents around the clock. According to Microsoft, MAI Code 2 reaches 96 percent of Claude Opus 5.5’s coding performance at about 5 percent of the price, while the same comparison gives it 98 percent of GPT-6.1 Sol’s performance at about 40 percent of the price.
MAI Code 2: Price becomes the competitive battleground
The coding claims are central to Microsoft’s argument that specialist models can undercut general-purpose frontier systems without sacrificing the performance required for common enterprise work. Microsoft AI CEO Mustafa Suleyman said the company’s commitment is to give you the very best price performance out there so you can really scale your agents.
Microsoft said MAI Code 2’s results are averages across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode. The company also reported that, in a blind front-end test, developers preferred MAI Code 2 over GPT-6 Sol 64 percent of the time. These are Microsoft’s own benchmark and preference results, so they establish a pricing strategy rather than independent proof of superiority.
The comparison has an important timing qualification. OpenAI released GPT-6.2 on the same day as Microsoft’s announcement, so Microsoft’s comparison with GPT-6.1 Sol already concerns an older generation, Decrypt reported. Microsoft also calculated MAI Code 2’s costs using a 50 percent promotional discount, which means the quoted price advantage may depend on the promotion.
A coding model with stated limits
Microsoft describes MAI Code 2 as a tool for everyday enterprise coding tasks, including the kind of work customers run in GitHub Copilot. The model and MAI Code 2 Flash are generally available in GitHub Copilot, while MAI Code 2 Flash is also in private preview in Microsoft Foundry. That gives developers immediate access to test Microsoft’s pricing claims against their own workloads.
The lighter Flash model targets high-volume tasks such as code explanations, targeted edits, routine fixes and rapid iterations. Microsoft says MAI Code 2 Flash matches GPT-6 Luna at 65 percent of its price and Claude Haiku 5.5 at 20 percent of the cost across three benchmarks for routine coding. Decrypt characterized those claims as evidence that the model could be extremely cheap without sacrificing quality, while noting that Anthropic already competes in the cheaper market with Haiku.
Microsoft also set a clear boundary on the coding models’ capabilities. The company said they prioritize the coding work its customers do every day and are explicitly not yet tuned for harder problems such as fundamental science research and very long horizon agentic work. Suleyman said Microsoft would not pursue those niche capabilities at customers’ expense because the added cost would apply to every task, including routine bug fixes.
Security is the immediate operational test
MAI Cyber 2 Flash extends the same specialization strategy into defensive security. Microsoft designed the model for continuous security monitoring and defensive work, including identifying and patching vulnerabilities in live production systems. The company says the model was built to deliver frontier intelligence at a fraction of the cost.
According to Microsoft, MAI Cyber 2 Flash matches GPT-6 Astra’s quality at less than 2 percent of the cost per task on the CyberGym-E2E benchmark. Its reported score was 92.3 percent, compared with 91.5 percent for GPT-6 Astra. Decrypt noted that CyberGym-E2E measures the percentage of software-security investigations completed successfully, and that Microsoft excluded 92 of 905 tasks after identifying overlap between the benchmark’s source material and its training data.
Usage data offers a measure of adoption. Microsoft said MAI Cyber accounts for almost 40 percent of scans in MDASH, while 90 percent of scans that started with MAI continued to use it. MAI Cyber 2 Flash is priced 86 percent lower on output tokens than MAI Cyber 1.1 Flash and is limited to Microsoft security products, with replacement for eligible MDASH customers scheduled to begin in mid-December.
The model has a 1 million token context window. Microsoft set its price at 0.10 dollars per million input tokens and 0.50 dollars per million output tokens. The combination of lower unit pricing, a large context window and continuous monitoring could make the model relevant for organizations that need automated checks to run on every code change, although the benchmark caveats remain important.
Image, speech and transcription widen the portfolio
The rest of the seven-model launch covers media and decision workloads. MAI Image 2.7 is in public preview on Foundry and through partners including fal, OpenRouter and Vercel. Artificial Analysis ranks it first for image editing, ahead of OpenAI’s GPT Image 2.6, and fourth for text-to-image generation. Decrypt’s testing found that its images beat GPT Image and Google’s Nano Banana on realism, creativity and overall quality, although those observations came from only a few generations.
MAI Realtime 1 is Microsoft’s first speech-to-speech model and starts at 0.01 dollars per audio minute. Microsoft said it outperformed GPT-Realtime 2.1 across external and Dynamics 365 benchmarks. Dragon Copilot Transcribe is based on MAI Transcribe 2 and handles about 10 million patient encounters each month. Microsoft said it halves error rates in multilingual workflows and is used by a third of US physicians.
Decision 1.1 is generally available and is a fast, cost-effective model designed to make reliable decisions in software at scale. It is priced at 0.042 dollars per million input tokens and adds model routing, allowing applications to reserve larger, more expensive models for tasks that require them. Decision models reach answers quickly within constrained definitions and provide calibrated confidence scores, allowing applications to defer when uncertainty is high.
Decision 1.1 builds on Microsoft-Decision-1, which was released in October. Microsoft said the new model delivers a significant improvement in intelligence and speed at a competitive price. It is built on a new base, Google’s Gemma 4 26B A4, adding multimodal support so it can analyze images in instructions.
Control and infrastructure support the pitch
Microsoft said each of the seven models is running in production on Maia 200, its second-generation AI accelerator. The company claims Maia 200 delivers up to 1.7 times more tokens per watt than leading GPU infrastructure. If sustained in customer deployments, greater token efficiency could reinforce the lower per-task pricing that underpins Microsoft’s specialist-model strategy.
MAI Licensed Weights extends the same approach to control. Microsoft says it lets organizations deploy and customize Microsoft AI models in their own environments on Azure through the Microsoft Frontier Company program. The first custom model Microsoft has made is for 5G network operation analytics, which the company says outperforms GPT-6.1 Sol at 50 times cheaper costs. Microsoft said AT&T developed the model and also found that Code 2 delivered frontier-level performance at a significantly lower price than alternatives on complex scenarios.
Microsoft is also updating Foundry IQ with an enterprise knowledge graph and human annotations in preview, allowing experts to rate and comment on agent traces. The company said those annotations feed an improvement loop combining optimization tests, evaluations and comparisons with real-world results. Microsoft Foundry has an expanded model pool featuring the top 50 models and offers more than 11,000 models from OpenAI, Anthropic, Meta, Mistral AI and the open source community.
The practical significance of MAI Code 2 is therefore not a single benchmark claim. Microsoft is presenting a portfolio in which cheaper specialists handle repeated tasks, licensed weights give enterprises more control, and Maia 200 supplies the infrastructure. The immediate test is whether GitHub Copilot users can reproduce the promised economics in production, while the mid-December MDASH rollout will show whether MAI Cyber 2 Flash can turn Microsoft’s benchmark claim into sustained security usage.

