State of AI for development: March 2026

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

GPT-5.4 raises the bar and lowers the price

OpenAI’s GPT-5.4 is its first unified coding and general model. It tied for first on the AA Intelligence Index, carries roughly a million tokens of context and scored 75 per cent on OSWorld for native computer use, above the human baseline. At $2.50 per million input tokens it is priced for everyday work, and GPT-5.4 Mini approaches the flagship’s coding quality at under a third of the cost. Both smaller variants are verbose with elevated hallucination rates.

What we are telling clients: evaluate GPT-5.4 now, especially for agentic tasks that touch the desktop. Route high-volume work to Mini where the economics matter, but keep review discipline tight on its output given the hallucination rates.

Agents can now see what they built

Computer use arrived in Claude Code and Cowork this month (macOS preview): the agent controls mouse, keyboard and screen, with a classifier-mediated approval mode. That closes the verify loop - the agent can run what it built, visually inspect the result, fix it and re-test. The gap between “the code works” and “the code does what I wanted” is where most bugs live, and until now agents could only read their way across it.

What we are telling clients: expect visual verification to become standard and start piloting it on internal tools where the blast radius is small. An agent with a mouse and keyboard needs the same access controls you would give a new starter, so treat the approval settings as seriously as the capability.

Infrastructure is being rebuilt for agents

Stripe shipped stripe projects add, which lets an agent provision services from the command line, and Ramp, Visa, ElevenLabs, Sendblue and Resend all shipped CLIs the same day. CLIs are increasingly preferred over MCP for agent tooling, and orchestration tools such as Cline Kanban now run multiple CLI agents in parallel across isolated worktrees.

What we are telling clients: if you build tools other developers use, assume agents are now part of your audience and ship a CLI. Internally, start treating agent access to your services as a design requirement, the way you once treated mobile.

---

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.