Deep|LLM: 26H1 Update (Part 1): Agent Scaling Law, Token Maxxing to ROI Discipline

1. Overview

The defining change in the AI sector in H1 2026 is not a step-up in any single model's raw capability, but that agent systems have begun to enter a phase of self-iteration. In our view, software engineering will be the first vertical to see AGI-level automation — driven not by sheer parameter scaling, but by a new Agent Scaling Law powered jointly by the agent harness, real-task feedback, and a synthetic-data flywheel.

The crux of this shift is that the scope of AI scaling is expanding from the model itself to the task environment, toolchain, and feedback systems in which the agent operates. Coding was the first scenario to work end-to-end, because it is inherently verifiable, testable, and reward-checkable, and most readily forms a synthetic-data flywheel. Going further, frontier model research and training are, broadly, coding tasks; frontier labs already use AI to accelerate them. If models go on to participate in their own research, training, and self-improvement, that could become an important marker of more general AGI.

At the model layer, H1 2026 remained driven by the Anthropic–OpenAI duopoly. Anthropic leads on coding agents, Claude Code enterprise penetration, and high-paying customers, while OpenAI is catching up quickly via GPT-5.5, Codex, its API, and the ChatGPT distribution system. While any single company's revenue growth may begin to slow from its steepest phase, taken together, Anthropic and OpenAI show model-layer growth that remains healthy — this should not be read simplistically as AI demand having peaked. Codex's evolution from a programmer's tool into an entry point for knowledge work also shows that the coding-agent feedback flywheel is spilling over into broader white-collar work.