Microsoft says MAI Code 2 nearly matches GPT-6.1 Sol at a fraction of the cost

At Ignite, Microsoft unveiled two in-house coding models: MAI Code 2 and MAI Code 2 Flash. Microsoft says MAI Code 2 comes close to GPT-6.1 Sol and Claude Opus 5.5 while costing only a fraction as much. Both models are available in GitHub Copilot starting today.

Model releases Nov 17, 2026 7 min read Source: Hacker News Startseite
Illustrative image, AI-generated

What's going on

On November 17, 2026, at its Ignite developer conference in San Francisco, Microsoft unveiled two new AI models for coding: MAI Code 2 and MAI Code 2 Flash. MAI stands for Microsoft AI, the AI division led by Mustafa Suleyman. Both models are available now in GitHub Copilot, the coding assistant from Microsoft subsidiary GitHub. MAI Code 2 Flash is also available as a limited-access preview in Microsoft Foundry, Microsoft's cloud platform for AI models. OpenRouter, Vercel, Hermes, OpenCode and Kilo are set to follow soon.

According to CNBC Arabia and CNN Business Arabic, citing Reuters, Microsoft announced more at Ignite: a cybersecurity model that only some customers get at first, plus new models for image generation and speech recognition. PANews (via KuCoin) says there were seven in-house MAI models in total. With them, Microsoft wants to compete with OpenAI and Anthropic, even though it's closely tied to OpenAI.

Two models, two jobs

Microsoft splits the work clearly. MAI Code 2 is meant for more demanding everyday enterprise tasks: getting code ready for production, thorough debugging, changes across multiple files, understanding how things fit together across a whole code repository, and building user interfaces. Microsoft promises changes that code owners would actually "merge," meaning accept into the project's main branch.

MAI Code 2 Flash is built for high volumes of smaller tasks: explaining code, targeted edits, routine fixes and fast iteration.

Microsoft is also upfront about the limits. Both models are "explicitly not yet" tuned for harder problems such as fundamental research or very long agentic tasks. Those are workflows where an AI works through many steps on its own for hours.

The numbers, according to Microsoft

Microsoft averaged four benchmarks, standardized test suites that reflect typical coding tasks in GitHub Copilot: DeepSWE 1.1, Terminal-Bench 2.1, SWE Atlas and FrontierCode v1.1. The exact figures come from Unite.AI's report:

Model Avg. success rate Avg. cost per task
MAI Code 2 63.2% $0.309
GPT-6.1 Sol 64.1% $0.761
Claude Opus 5.5 65.3% $6.306

That's the basis for Microsoft's main claim: MAI Code 2 delivers 98 percent of GPT-6.1 Sol's performance at 40 percent of the cost, and 96 percent of Claude Opus 5.5's performance at just 5 percent of the cost. So the models are close on quality and far apart on cost.

For the small models, Microsoft gives three benchmarks (without FrontierCode):

Model Avg. success rate Avg. cost per task
MAI Code 2 Flash 61.0% $0.052
GPT-6 Luna 55.1% $0.080
Claude Haiku 5.5 53.9% $0.319

Microsoft's blog post only claims "comparable performance." Going by the numbers, though, Flash actually comes out ahead, at 65 and 20 percent of the cost respectively.

There's also a visual coding test. People blindly compared output from MAI Code 2 and GPT-6 Sol on user interfaces, games, data visualizations, 3D scenes and SVG graphics, judging both looks and functionality. 258 reviewers took part in total. Microsoft says they preferred MAI Code 2 64 percent of the time. Unite.AI, however, reports 67 percent. It's unclear where the difference comes from.

Crypto Briefing also reports on tests at telecom giant AT&T. There, MAI Code 2 Flash reportedly beat certain GPT-6.1 variants while costing 50 times less. Which variants and which tasks isn't specified.

The fine print

The footnotes matter for putting this in context:

  • All models ran at maximum reasoning effort, meaning the highest level of "thinking time" before answering.
  • Three of the four benchmarks ran on MSBench, Microsoft's internal testing platform, via the GitHub Copilot CLI and without subagents (helper AIs that subtasks get handed off to). Only the FrontierCode scores come from independent analytics firm Artificial Analysis.
  • Costs for MAI Code 2 and Code 2 Flash are based on a promotional price with a 50 percent discount. At the regular price, the cost advantage would be roughly half as big. Against Claude Opus 5.5 it would still be large, but against GPT-6.1 Sol it would almost vanish: roughly $0.62 instead of $0.31, versus $0.76.
  • Microsoft itself notes that the comparisons can shift depending on workload, configuration, reasoning effort and pricing assumptions.

Crypto Briefing stresses that independent benchmarks haven't confirmed the performance claims yet. Vendor tests are "a starting point, not a verdict."

From coding model to office workhorse

Microsoft has derived another model from Code 2 Flash: MAI Work Flash. It's meant to work across several Microsoft 365 products and Fabric IQ. The figures come from internal tests:

  • Excel and PowerPoint: on par with GPT-6 Luna in accuracy and quality respectively, and in token efficiency. Tokens are the chunks of text that AI usage is billed by.
  • Word: similar performance to GPT-6.1 Sol across five internal benchmarks, with a "significantly smaller" model and about 42 percent fewer input tokens per call.
  • Teams: better than GPT-6 Luna at tracking decisions, filling knowledge gaps and following its own instructions, with up to 28 percent fewer tokens.
  • Fabric IQ / Power BI: For generating DAX queries, MAI Work Flash scores 0.77 overall according to Unite.AI, versus 0.54 for GPT-6 Astra, 0.45 for GPT-5.6 Sol and 0.30 for Opus 5.5. DAX is Power BI's query language. This capability powers the "Chat with your Data" feature, which lets you ask questions about business data in plain language.

According to PANews, Microsoft also offers companies the option to license the model weights so they can customize the models and run them on Azure. You can't download them freely. For customization, Microsoft points to its "Custom Models" page.

How we got here

According to Crypto Briefing, MAI Code 2 is already the third generation of Microsoft's in-house coding models. MAI Code 1 Flash, the first in-house coding model for Copilot, came out in June 2026. MAI Code 1.1 Flash followed in August. It could also process images and was about 73 percent cheaper. Both models reportedly led developers to accept more suggestions, and they used fewer tokens.

Mustafa Suleyman told Reuters that the strategy of specialized models for individual domains keeps costs low and makes sure Microsoft keeps pace with the competition, or even beats it on some benchmarks:

"This isn't just a single model, and it's not a one-off. It's a system, a kind of machine designed to keep improving and find the best balance between performance and price."

Suleyman also pointed to recent security incidents, such as the break-in by OpenAI agents into the Hugging Face platform mentioned in the reports. That's why, alongside performance, he says he also prioritizes "alignment," meaning that an AI doesn't take harmful actions and behaves as intended. Fittingly, PANews mentions the security model MAI Cyber 2 Flash: on the CyberGymE2E benchmark, it's said to match GPT-6 Astra's quality at less than 2 percent of the cost per task.

Our take

Microsoft isn't pitching the strongest model. It's pitching the best price-performance. In Microsoft's own numbers, MAI Code 2 trails GPT-6.1 Sol and Claude Opus 5.5 slightly. But for companies whose AI bills grow with every developer they add, a much cheaper model with nearly the same quality can still be the more attractive option. Strategically, this makes Microsoft less dependent on OpenAI and Anthropic, whose models have powered much of Copilot so far. Whoever owns the model also controls the costs.

There are three reasons for caution. First, almost all the numbers come from Microsoft's own tests. Second, the 50 percent promotional price flatters the cost comparison, and nobody knows how long it will last. Third, Microsoft itself names a limit: MAI Code 2 isn't built yet for long, autonomous agentic tasks, which are among the most important use cases for top-tier models right now. For Copilot users, that means it may be worth trying for routine work. For really hard problems, the pricier top models remain the safer choice for now. Independent tests are still to come.

Related videos

Microsoft Ignite: Opening Keynote 2026 Microsoft · YouTube

Sources

#microsoft#coding#github copilot#mai#ignite
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.