State of AI for development: January 2026

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

MCP became the enterprise integration standard

Microsoft shipped production-ready MCP support for Azure Functions with built-in authentication. Salesforce added beta MCP support to Agentforce with governance covering more than 10,000 public MCP servers, and SAP shipped an MCP server for Commerce Cloud. Each vendor moved independently in the same month. Anthropic also launched the MCP Apps specification, which lets servers return interactive UI components rather than plain text.

What we are telling clients: December’s advice to build connectors as MCP servers still holds; the new work is governance. Decide which servers your organisation may use before the sprawl arrives - Salesforce is already treating public servers like an app store with controls.

An open-weight model arrived at a tenth of the price

Moonshot’s Kimi K2.5, a trillion-parameter mixture-of-experts model with open weights, claims open-model state of the art on coding (76.8 per cent on SWE-bench) at API pricing around 10 per cent of Opus. It does not quite match Opus for feel, but it is remarkably close on the benchmarks that matter for agentic coding.

What we are telling clients: evaluate it now for cost-sensitive workloads. A model at a tenth of the cost and close to frontier quality changes the economics of high-volume agentic work, and a proven second option reduces your dependence on any single provider.

Context engineering now beats model choice

The most actionable theme of the month came from several independent directions. Cursor cut token usage by 47 per cent through dynamic context management. Practitioners converged on a one-session-per-task workflow for Claude Code because performance degrades once context passes about 40 per cent of capacity, and the quadratic cost of attention means that degradation is non-linear.

What we are telling clients: invest engineering time in what goes into an agent’s context and when to clear it, ahead of debating which model to use. Disciplined context management is currently a larger and cheaper source of gains than switching models.

---

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.