MAI Code 2, Sashiko, Claude Code

Industry

Dev Tools

Overview

Microsoft's MAI Code 2 launch received coverage from Axios, ZDNET, and InfoWorld as the company combines lower-cost models with GitHub, Visual Studio, and Azure distribution. AWS and Google Cloud shifted attention from code generation to review and trust: high-AI-adoption teams merged 98% more pull requests while review time rose 91%, and Google's Sashiko completed roughly 241,000 Linux kernel reviews. Enterprise economics are becoming visible, with top Claude Code users averaging about $1,500 per month and Databricks identifying $1.2 million in annual waste from seven MCP bugs. The market is converging on measurable workflow outcomes, governance, and integration rather than model capability alone.

Key Stories

Microsoft's MAI Code 2 launch is a direct challenge to external model providers. Axios framed the strategy around specialized models and lower prices, while ZDNET reported Microsoft's claim that Code 2 delivers 96% of Claude Opus 5.5's performance at 5% of the cost. InfoWorld added benchmark results of 63.2% for Code 2 and 61% for Code 2 Flash, but noted that integration, security, testing, and review costs are excluded. The development gives Microsoft a way to reduce dependence on OpenAI and Anthropic while using its existing developer distribution. ZDNET also covered the launch.

AWS and Google Cloud independently identified code review as the constraint on AI-assisted development. Amazon Web Services and Google Cloud Blog reported the findings. AWS cited research showing 98% more merged pull requests but a 91% increase in review time, while Google's Sashiko completed about 1,200 reviews daily and coincided with Linux's 60-day follow-up fix rate falling from 3.4% to 2.2%. The implication is clear: agentic coding products need evidence, policy controls, and human approval built into the workflow, not just stronger generation.

Claude Code has become a material operating expense. Business Insider reported some projects generating tens of thousands of dollars in bills and Anthropic reporting 11,900 business customers spending at least $100,000 in the prior year. Databricks showed the countervailing opportunity, estimating $1.2 million in annual savings from seven MCP-server bugs and 12,000 engineer hours. Cost controls, routing, trace analysis, and usage governance are becoming product requirements as adoption scales.

Google launched Gemini 4 Neon for multi-file debugging, refactoring, and parallel subagents at $3 per million input tokens and $15 per million output tokens. InfoWorld reported the model. The article still described supervised autonomy as the near-term operating model because production reliability and developer accountability remain unresolved. OpenAI's own research similarly found coding-agent success on two-to-four-hour tasks rose from 23% in March to 46% in October, while human judgment remained necessary OpenAI.

Developer tools are expanding beyond conventional software teams. TechCrunch covered Snapchat's Easy Lens CLI, which supports more than 100 AR building blocks and 50 AI asset generators; IEEE Spectrum described Unreal Engine 6's planned shift from Blueprints to Verse and Scene Graph; and WIRED reported Lovable's move toward automated support for small-business operations. These examples widen the addressable market but also increase exposure to security, permissions, and quality-control failures.

See your coverage on Delve. See your coverage on Delve.