Microsoft Introduces Task-Specific AI Models for Coding and Security

Microsoft introduced a lineup of specialized artificial intelligence models at its Ignite conference on Tuesday, targeting software development, cybersecurity and image creation. The company says its models can handle certain tasks at lower cost than competing systems, although several of its comparisons depend on particular tests and pricing assumptions.

Coding models and their benchmarks

Microsoft says MAI Code 2 achieved 96% of Claude Opus 5.5’s performance on a group of enterprise coding evaluations at 5% of the cost. It also claims the model reached 98% of GPT-6.1 Sol’s performance for 40% of its price. OpenAI released a newer model, GPT-6.2, on the same day as Microsoft’s announcement.

The coding comparison averages four evaluations involving tasks such as repairing software and making changes across multiple files. Microsoft conducted three tests on its own evaluation platform and used external results for the fourth. Its cost calculation for MAI Code 2 included a 50% promotional discount.

A smaller model, Code 2 Flash, is intended for routine edits. Microsoft says it performs similarly to Claude Haiku 5.5 at one-fifth of the price. Both coding models are available through GitHub Copilot.

Security and image tools

Microsoft says MAI Cyber 2 Flash completed 92.3% of the vulnerability-discovery tasks it evaluated on CyberGym-E2E, compared with 91.5% for GPT-6 Astra, while costing less than 2% as much per task. Microsoft removed 92 of the benchmark’s 905 tasks because their source material overlapped with the model’s training data. The company has not published a direct comparison with OpenAI’s restricted GPT-6 Cyber model.

The security model is scheduled to begin replacing an older MAI model for eligible MDASH customers in mid-December. Microsoft also introduced MAI Image 2.7, which leads an Artificial Analysis image-editing ranking by a narrow margin; its uncertainty range overlaps with that of the second-ranked model.

Why it matters

For someone learning about crypto, this is an AI announcement, not a change to a cryptocurrency or blockchain. It does illustrate a useful principle for assessing technology claims: a tool designed for one job can be less costly on that job without being better at everything. Benchmark choices, excluded tests and temporary discounts all affect how comparisons should be understood.

Originally reported by Decrypt.