cd /news/ai-products/microsoft-launches-mai-code-2-models… · home › topics › ai-products › article
[ARTICLE · art-171759] src=infoworld.com ↗ pub= topic=ai-products verified=true sentiment=↑ positive

Microsoft launches MAI Code 2 models to challenge OpenAI and Anthropic in AI coding

Microsoft launched two in-house coding models, MAI Code 2 and MAI Code 2 Flash, at Ignite 2026, positioning them against OpenAI and Anthropic offerings. Microsoft reported MAI Code 2 averaged a 63.2% pass rate across DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1, versus 64.1% for OpenAI's GPT-6.1 Sol and 65.3% for Anthropic's Claude Opus 5.5, while MAI Code 2 Flash averaged 61% against GPT-6 Luna's 55.1% and Claude Haiku 5.5's 53.9%. Microsoft said average cost per task was $0.309 for MAI Code 2, $0.761 for GPT-6.1 Sol and $6.306 for Claude Opus 5.5, with MAI Code 2 priced at a promotional $0.50 per million input tokens and $2.50 per million output tokens until May 17, 2027.

key takeaway

MAI Code 2 averages a 63.2% pass rate across four GitHub Copilot coding benchmarks at $0.309 per task, undercutting GPT-6.1 Sol ($0.761) and Claude Opus 5.5 ($6.306) on cost while trailing them slightly on average score.

read4 min views12 publishedNov 18, 2026

Microsoft has launched a pair of AI coding models that it says can compete with offerings from OpenAI and Anthropic, marking its latest effort to build proprietary AI capabilities and reduce reliance on the partners that helped fuel its generative AI push.

The two new models, MAI Code 2 and MAI Code 2 Flash, were launched at Ignite 2026. MAI Code 2 is designed for common enterprise coding tasks such as deep debugging, implementing changes across multiple files, repository-wide reasoning, and user interface development. MAI Code 2 Flash, meanwhile, targets high-volume tasks such as code explanations, targeted edits, routine fixes and rapid iterations.

Microsoft said it built these models to prioritize the coding work its customers do every day. The new models are explicitly not yet tuned for harder problems, like fundamental science research and long-horizon agentic work, the company noted in the press statement.

The launch also marks another step in Microsoft’s effort to build its own AI model capabilities and reduce its reliance on external providers.

With the launch of new coding models, Microsoft reduces its dependence on OpenAI and Anthropic, said Pareekh Jain, CEO at EIIRTrend & Pareekh Consulting. “Microsoft also has a major advantage through GitHub, Visual Studio and Azure, which are already used by millions of developers. Its immediate goal is to strengthen its developer ecosystem and reduce AI costs. So it doesn’t have to pay OpenAI and Anthropic every time someone uses GitHub Copilot. Microsoft is protecting its profits and locking developers into its own tools,” he added.

Microsoft compared MAI Code 2 with competing models across four widely used benchmarks in GitHub Copilot: DeepSWE 1.1, Terminal-Bench 2.1, SWE-Atlas and FrontierCode v1.1.

In the average pass rate across these benchmarks, MAI Code 2 recorded an average score of 63.2%, compared with 64.1% for OpenAI’s GPT-6.1 Sol and 65.3% for Anthropic’s Claude Opus 5.5.

On individual benchmarks, MAI Code 2 scored 73.2% on DeepSWE 1.1, 84.6% on Terminal-Bench 2.1, 43.1% on SWE-Atlas, and 52% on FrontierCode v1.1. In comparison, GPT-6.1 Sol scored 74.6%, 85%, 47.4%, and 49.4%, respectively, while Claude Opus 5.5 scored 69%, 78.7%, 57.5%, and 56%.

For MAI Code 2 Flash, Microsoft has compared it with GPT-6 Luna and Claude Haiku 5.5 across DeepSWE 1.1, Terminal-Bench 2.1, and SWE Atlas benchmarks at maximum reasoning. In the average pass rate across the benchmarks, MAI Code 2 Flash scored 61%, ahead of GPT-6 Luna at 55.1% and Claude Haiku 5.5 at 53.9%. On individual benchmarks, MAI Code 2 Flash recorded 59.9% on DeepSWE 1.1, 78.3% on Terminal-Bench 2.1, and 44.8% on SWE Atlas. GPT-6 Luna scored 59.6%, 69.3%, and 36.4%, respectively, while Claude Haiku 5.5 scored 46%, 73.4%, and 42.1%, respectively.

Microsoft claimed the average cost per task came out to $0.309 for MAI Code 2, for GPT-6.1 Sol at $0.761, and Claude Opus 5.5 at $6.306. Similarly, for MAI Code 2 Flash, the average cost per task was $0.052 as compared to $0.080 for GPT-6 Luna and $0.319 for Claude Haiku 5.5. However, MAI-Code-2 and Flash costs, which have been used for comparison, are at 50% promotional pricing.

According to GitHub’s pricing page, MAI Code 2 is available at the promotional pricing of $0.50 per million input tokens, $0.025 per million cached input tokens, and $2.50 per million output tokens until May 17, 2027. Post this, the pricing will be revised to $1.00 per million input tokens, $0.05 per million cached input tokens, and $5.00 per million output tokens.

Similarly, the promotional pricing for MAI Code 2 Flash is $0.05 per million input tokens, $0.005 per million cached input tokens, and $0.25 per million output tokens. This will be increased to $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.50 per million output tokens.

“Microsoft’s own savings are measured at a 50% promotional price, so the gap narrows sharply once the promotion ends. And the cost it compares is the spend of a benchmark run, not a published price per token,” said Sanchit Vir Gogia, chief analyst at Greyhound Research.

While Microsoft’s benchmark results highlight the models’ coding capabilities and cost advantages at promotional pricing, enterprises need to look beyond model costs, as these account for only part of the overall expense. “Enterprises must also pay for integration, testing, security, human reviews and fixing AI-generated errors,” said Jain. He explained that a 60% reduction in model costs would deliver only 6% in overall savings if model usage accounts for 10% of total development costs. The real benefit comes from reducing the total cost of delivering working software, not just generating cheaper code.

Gogia cautioned that before adopting either model, technology leaders should test it on their own code, through their own tools, at the effort setting they will use in production. They should then track the cost per merged code change, change failure rates, and security findings against a baseline from teams not using AI coding models.

He also pointed to Google’s DORA research, which has associated wider AI adoption with weaker delivery stability. This is where hidden costs can emerge. Security is another concern. Microsoft’s model cards rate both models highly on offensive cyber capabilities. Gogia said CIOs should therefore ensure that pilot programmes assess not just the quality of AI-generated code but also its security and reliability when integrated into production systems.

── more in #ai-products 4 stories · sorted by recency
── more on @microsoft 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/microsoft-launches-m…] indexed:0 read:4min 2026-11-18 · —