State of AI for development: December 2025

Three things from December that we think change decisions, and what we are telling clients to do about each.

MCP is now the safe bet for agent integration

Anthropic transferred the Model Context Protocol to the Linux Foundation this month, under a new Agentic AI Foundation, with OpenAI, Google, Microsoft and AWS all backing it. A protocol that began as one vendor’s design is now an open standard with every major provider behind it.

What we are telling clients: if you are connecting agents to internal systems, build those connectors as MCP servers. The risk of that work being stranded by a format war has largely gone, and the same connector should serve whichever models you run against it.

The skills pattern is spreading beyond one vendor

Skills, the markdown-plus-scripts pattern we flagged in October as worth adopting early, arrived in the Claude app this month, and OpenAI’s Codex adopted something similar. The picture emerging is that MCP handles how agents connect to tools while skills capture how you want the work done.

What we are telling clients: start encoding your team’s working procedures and domain knowledge as skills. Six months ago this looked like an investment in a single tool; it now looks portable, and teams that write this material down get more consistent results from whichever agent runs it.

Capable models got dramatically cheaper

Google’s Gemini 3.0 Flash outperforms Gemini 3 Pro on some agentic coding benchmarks despite sitting in the cheaper tier, because it received more refined reinforcement learning, and its $0.50/$3 pricing is remarkable. DeepSeek V3.2 offers GPT-5-tier capability with open weights at roughly 6 per cent of the cost.

What we are telling clients: revisit any model choice made more than a quarter ago. Reserve the expensive frontier models for tasks where cheaper ones demonstrably fail, and route cost-sensitive or high-volume work to the fast tiers, which are now genuinely capable. Ignore the weekly benchmark records; the difference between 75 and 78 per cent on SWE-bench does not translate to a meaningfully different development experience.

---

This digest is the public edition. Advisory clients get the full monthly review, a tailored edition for their context and the conversation that goes with it.