AI token spending jumped from under $50 in a year to $1,000 in a single day after coding agents became broadly useful in 2026, Simon Willison said in a closing keynote at the WeAreDevelopers World Congress North America in San Jose. The surge forced companies to tighten token and spend rules as coding agents achieved clear product-market fit.
Willison's talk, delivered on Friday, traced the chronological arc of large language model progress through 2026. He pointed to November 2025 as the real starting point, when two model releases crossed what he called "an invisible line where something that didn't really work starts working."
The models that changed the equation
Claude Opus 4.5 and GPT-5.1 arrived in November 2025. On their own, Willison said, these were incremental improvements on previous models. But paired with existing coding agent tools - Claude Code, which launched in February 2025, and the younger Codex - they produced a reliability leap. The agents stopped being experimental curiosities and started delivering consistent, practical results for developers.
The spending data told the story. Willison noted that his own AI token costs had stayed below $50 over the preceding year. After the coding agents clicked, he burned through $1,000 in one day. That spending pattern repeated across teams and organizations, triggering a new wave of enterprise adoption for tools like Claw, a software platform that gained traction alongside the agent improvements.
From product-market fit to cost controls
The rapid adoption created a predictable second-order effect. Companies that had encouraged experimentation suddenly faced ballooning API bills. Willison described a shift toward tightening token and spend rules - not because the tools lost value, but because their utility had scaled faster than governance. Organizations began setting usage limits, tracking per-project costs, and reevaluating which workflows justified agent-level compute.
For developers and technical leaders who want to build these skills, AI Code Generation Courses cover the practical side of working with coding agents and generative code tools. The shift also intersects with broader automation strategies - AI Agent Courses address how autonomous systems fit into existing workflows.
Why this matters for executives and strategy leaders
The 2026 pattern reveals a cycle that repeats across AI adoption: capability crosses a threshold, usage explodes, and cost controls follow. For leaders in IT, HR, and operations, the signal is clear. Budgeting for AI tools cannot be static. When a model improvement makes a previously unreliable task suddenly reliable, spending can spike 20x or more in days - not quarters. Teams that do not have spend governance in place before that threshold arrives will face hard conversations with finance after the fact.
Your membership also unlocks:
