{"slug": "microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the", "title": "Microsoft says its new MAI Code 2 does 96% of Claude Opus 5.5’s work for 5% of the price", "summary": "Microsoft unveiled seven MAI models at its Ignite conference in San Francisco on Tuesday, led by the coding model MAI Code 2, which Microsoft says reaches 96% of Claude Opus 5.5's quality at 5% of the cost and is already generally available in GitHub Copilot. Microsoft's own figures put MAI Code 2 at 63.2% average across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, against 64.1% for GPT-6.1 Sol and 65.3% for Opus 5.5, at a promotional $0.309 per task versus $0.761 and $6.306. The comparisons rely on Microsoft-chosen benchmarks and a 50% discount, and Microsoft says the coding models are \"explicitly not yet tuned\" for fundamental science research and very long agentic tasks.", "body_md": "Microsoft’s launch artwork for MAI Code 2 and MAI Code 2 Flash. Image: Image: Microsoft\n\nMicrosoft [unveiled a family of seven MAI models](https://microsoft.ai/news/microsoft-ignite-2026/) at its Ignite conference in San Francisco on Tuesday, taking on OpenAI and Anthropic with its own AI on price. Its flagship coding model, MAI Code 2, scores 96% of Claude Opus 5.5’s result at 5% of the cost, Microsoft says, and it is already in GitHub Copilot.\n\n## What Microsoft launched\n\nThe line-up covers coding, security, images and voice. These are Microsoft’s own figures:\n\n- **MAI Code 2:** 98% of GPT-6.1 Sol’s coding quality at about 40% of the price, and 96% of Claude Opus 5.5’s at 5%, averaged over four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1).\n- **MAI Code 2 Flash:** comparable to GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20%, for quick edits, routine fixes and code explanations.\n- **MAI Cyber 2 Flash:** 92.3% on the CyberGym-E2E vulnerability-finding benchmark, against 91.5% for GPT-6 Astra, at less than 2% of the cost per task, with a one-million-token context window.\n- **MAI Image 2.7:** first on Artificial Analysis’s image editing leaderboard, ahead of GPT Image 2.6 Sunburst (max) at a fifth of the price.\n- **MAI Realtime 1:** Microsoft’s first live speech-to-speech model, from $0.01 per audio minute, which it says beats GPT-Realtime 2.1.\n\nThe seven also include Dragon Copilot Transcribe, built on MAI Transcribe 2, which Microsoft says now transcribes every doctor and patient conversation in its Dragon Copilot clinical tool, and Decision 1.1, the latest version of the routing model it [launched as Microsoft-Decision-1](https://madrobot.blog/2026/10/11/microsoft-decision-1-ai-model-alibaba-qwen-foundry/) in October, which sends work to bigger models only when needed.\n\n## Where you can use them\n\nMAI Code 2 and Code 2 Flash are generally available in GitHub Copilot, in the model picker and through Copilot Fusion, chief executive Satya Nadella said in his [opening keynote](https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/). Code 2 Flash is also in a limited-access preview in Microsoft Foundry, where MAI Image 2.7 and a cheaper Image 2.7 Flash are in public preview at about $35 and $10 per 1,000 images.\n\nCyber 2 Flash is going only into Microsoft’s own security products: it replaces the existing model in the MAI profile of its MDASH scanning tool for eligible customers from mid-December. Realtime 1 is in preview through MAI Labs. Microsoft hasn’t given a country-by-country list for any of the models.\n\n## Read the small print\n\nEvery comparison is Microsoft’s own, on benchmarks it picked, and the coding prices lean on a discount. The [MAI Code 2 post](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/) shows its cost at a 50% promotional price: $0.309 a task on average, against $0.761 for GPT-6.1 Sol and $6.306 for Opus 5.5, while scoring 63.2% to their 64.1% and 65.3%. At full price it would cost about four-fifths as much as Sol and a tenth as much as Opus. Code 2 Flash’s figures use the same discount.\n\nMicrosoft says the coding models are “explicitly not yet tuned” for harder work such as fundamental science research and very long agentic tasks. MAI Image 2.7’s lead is 4 Elo points, as of November 16. For [Cyber 2 Flash](https://microsoft.ai/news/introducing-mai-cyber-2-flash/), Microsoft found that 92 of CyberGym-E2E’s 905 tasks drew on the same open source projects and vulnerability reports as some of its training data, so it dropped them for every model. It hasn’t yet tested OpenAI’s GPT-6 Cyber, which only select customers can use.\n\nThe targets are moving, too. OpenAI [launched GPT-6.2 Astra and Sol](https://madrobot.blog/2026/11/17/openai-gpt-6-2-astra-sol-price-benchmarks/) the same day, cutting Astra’s price by 20%, and Nadella said GPT-6.2 is coming to Foundry.\n\n## ‘My team is building true superintelligence’\n\nMustafa Suleyman, chief executive of Microsoft AI, told [Axios](https://www.axios.com/2026/11/17/microsoft-ai-models-openai-anthropic-ignite) that specialising lets Microsoft “cram more tokens and knowledge and expertise into the same amount of parameters”. He said it has reached state-of-the-art pre-training results and is making strong progress on a frontier general-purpose model, building on the [2.8 trillion parameter MAI-Base-2](https://madrobot.blog/2026/10/16/microsoft-mai-base-2-2-8-trillion-parameter-ai-model-deepseek/) it revealed in October, but gave no release date. He was blunt about the aim:\n\nMy team is building true superintelligence.\n\nMustafa Suleyman, chief executive of Microsoft AI, to Axios\n\nMicrosoft will also let trusted business customers license MAI model weights, adapt them with their own data and run them in their own Azure environment, through its Frontier Company programme “soon”. It says [AT&T’s customised model](https://microsoft.ai/news/introducing-mai-licensed-weights/) beat GPT-6.1 Sol on 5G network analysis at a fiftieth of the cost. Suleyman told Axios the approach would “feel like open source”, though the models can run only within Azure. He declined to say how much GitHub Copilot traffic Microsoft expects to move to its own models, but predicted significant adoption.\n\n## Why it matters\n\nMicrosoft still sells OpenAI’s and Anthropic’s models through GitHub Copilot and Foundry, but every task that moves to an MAI model is one it no longer pays a partner for. If independent tests back up the numbers, cheap specialised models could put pressure on what its rivals charge for everyday work.\n\n*Sources: [Microsoft AI](https://microsoft.ai/news/microsoft-ignite-2026/) (announcement), with posts on [MAI Code 2](https://microsoft.ai/news/introducing-mai-code-2-and-mai-code-2-flash/), [Cyber 2 Flash](https://microsoft.ai/news/introducing-mai-cyber-2-flash/), [Image 2.7](https://microsoft.ai/news/introducing-mai-image-2-7/) and [Licensed Weights](https://microsoft.ai/news/introducing-mai-licensed-weights/); [Official Microsoft Blog](https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/) (Nadella keynote); [Axios](https://www.axios.com/2026/11/17/microsoft-ai-models-openai-anthropic-ignite) (Suleyman interview).*", "url": "https://wpnews.pro/news/microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the", "canonical_source": "https://madrobot.blog/2026/11/17/microsoft-mai-code-2-ignite-cheaper-openai-anthropic/", "published_at": "2026-11-17 19:00:00+00:00", "updated_at": "2026-11-17 19:53:01.002679+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-tools", "generative-ai"], "entities": ["Microsoft", "MAI Code 2", "MAI Code 2 Flash", "Claude Opus 5.5", "GPT-6.1 Sol", "GitHub Copilot", "Satya Nadella", "Mustafa Suleyman"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the", "markdown": "https://wpnews.pro/news/microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the.md", "text": "https://wpnews.pro/news/microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the.txt", "jsonld": "https://wpnews.pro/news/microsoft-says-its-new-mai-code-2-does-96-of-claude-opus-5-5s-work-for-5-of-the.jsonld"}}