State of AI for development: April 2026

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

The harness now matters more than the model

April turned “invest in the harness” from a contested claim into a working consensus. An 8B open model scored 33 out of 507 on a long-horizon reasoning benchmark with a good scaffold where the bare model scored zero. A harness-engineering paper lifted Terminal-Bench 2 pass rates from 69.7 per cent to 77 per cent in ten iterations, beating the human-designed baseline it started from. The scaffold and the evaluation loop are doing more of the lifting than the model.

What we are telling clients: put engineering effort into your harness (instruction files and evaluation loops) before chasing new model releases. The gains are larger and they carry over when you switch models.

Agent economics are about to become visible

GitHub Copilot announced a ninefold price increase for Claude models and a move to usage-based billing from 1 June. That is one point in a wider shift: Sam Altman says OpenAI has to become an inference company, and Together AI’s volume grew from 30 billion to 300 trillion tokens a month in a year. The unlimited-feeling subscription tier is ending for serious agent workloads.

What we are telling clients: get agent spend measured now, before June’s billing changes land, and adopt the cheap-executor-plus-expensive-advisor pattern. Pairing a small model with a frontier model for the hard judgments more than doubled one agentic benchmark score against the small model alone, and it preserves quality on the hard 10 per cent of decisions at a fraction of the cost.

Tool usage is up 65 per cent; delivery is up 8 per cent

The DX longitudinal study across more than 400 engineering organisations landed in April: AI tool usage rose 65 per cent while median pull-request throughput rose roughly 8 per cent. Coding is only about 14 per cent of a developer’s job, and AI is introducing new bottlenecks in code review and technical debt. The same conversation dominated AI Engineer Miami, where the argument was for depth (long serial loops on hard problems) over fifty parallel runs.

What we are telling clients: measure delivery outcomes rather than token consumption, and reward depth over breadth. Then attack the bottlenecks the data points at: review capacity and accumulating technical debt.

---

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.