Newshot

Microsoft's Most Cost-Effective Coding AI Model: MAI Code 2 Costs Just 5% of Claude Opus 5.5 Per Task

AI summary

At Ignite 2026, Microsoft announced its in-house coding models MAI Code 2 and MAI Code 2 Flash, now available in GitHub Copilot. Positioned for complex software development and high-frequency everyday coding respectively, the two models aim to strike a better balance between code quality and cost. On average across four published coding benchmarks, Code 2 reaches about 98% of GPT-6.1 Sol's performance at roughly 40% of its price, with an average cost per task of just $0.309—far below GPT-6.1 Sol's $0.761 and only about 5% of Claude Opus 5.5's $6.306.

Translation(model translation, not the source's own text)

Microsoft's Best Value Coding AI Model: MAI Code 2 Hits ~96% of Claude Opus 5.5's Performance at 5% of the Cost

IT Home, November 18 news: At today's Ignite 2026 event, Microsoft released its in-house coding models MAI Code 2 and MAI Code 2 Flash, and launched them on GitHub Copilot.

IT Home attaches the related video below:

In terms of positioning, the two products target complex software development and high-frequency everyday coding respectively, aiming to strike a better balance between code quality and usage cost.

The Copilot model picker showing MAI Code 2 and MAI Code 2 Flash
The Copilot model picker showing MAI Code 2 and MAI Code 2 Flash

Among them, Code 2 is positioned as a general-purpose coding model, usable for deep troubleshooting of program errors, modifying code across multiple files, understanding an entire code repository, and developing user interfaces. Microsoft says the model emphasizes directly delivering practical code change proposals, reducing unnecessary back-and-forth between developers and AI, and helping code move toward actual deployment.

Code 2 Flash focuses more on lightweight, high-frequency work, such as explaining code, making targeted changes, fixing common problems, and rapid iteration. Both models support a working style in which the AI invokes tools and advances development tasks step by step.

Value for money is the main selling point of this launch. Microsoft says that in the averaged results of the four coding benchmarks it published, Code 2 reaches about 98% of GPT-6.1 Sol's performance at roughly 40% of its price.

Official test data shows that MAI Code 2's average cost per task is only $0.309, far below GPT-6.1 Sol's $0.761, and just about 5% of Claude Opus 5.5's ($6.306). While maintaining top-tier engineering performance, it helps enterprises sharply cut their AI compute bills.

The lightweight, high-efficiency MAI Code 2 Flash, under a 50% promotional price, has an average cost per task as low as $0.052. Its performance matches or even exceeds competitors' flagship lightweight models, while its cost is only 65% of GPT-6 Luna's and 20% of Claude Haiku 5.5's.

In the three authoritative tests DeepSWE 1.1, Terminal-Bench 2.1, and SWE Atlas, MAI Code 2 Flash achieved an average pass rate of 61.0%, surpassing GPT-6 Luna (55.1%) and Claude Haiku 5.5 (53.9%). Delivering higher code delivery quality at a lower price, it redefines the value benchmark for lightweight coding models.

In terms of access thresholds, Code 2 is available to eligible subscribers such as Copilot Pro, Business, and Enterprise; Flash's coverage also includes Free and Student users. Enterprise developers can choose different models based on task difficulty and invocation scale.

Related Articles

  • Microsoft Win11 26H2 memory optimization hands-on: 8GB memory machines see noticeably improved experience
  • AI shakes up university math classrooms, Caltech professor no longer grades homework based on whether answers are right or wrong
  • Beating traditional K-pop idols, an AI-assisted song tops Korea's popular music charts for the first time
  • Marsh: There are signs hackers are using AI tools and agents to find security vulnerabilities
  • European cloud service company OVH enters AI models, open-sources Olala mini 7A1B
  • Google Gemini 4 Argon safety test out of control: 40 AI agents attack real websites on their own and share vulnerabilities with each other

The translation is machine-generated and individual terms may be off. Nothing is rewritten or cut; the original is authoritative.

Published by IT之家
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗