The biggest capability jump since GPT-4 didn't come from a bigger model. It came from letting models think — and training them, by trial and error, to think well.
The recipe
Take problems with checkable answers — math, code that must pass tests. Let the model generate long chains of thought; reward the ones that reach correct answers (reinforcement learning on verifiable rewards). The startling finding, shown openly by DeepSeek's R1-Zero: reflection, backtracking and self-correction emerge from this loop without being taught. o1 proved it first behind closed doors; within a year every frontier lab shipped a thinking mode.
The new scaling law
Test-time compute: accuracy rises with tokens spent thinking — a dial from instant-and-cheap to slow-and-brilliant on the same weights. AIME math went from ~12% (GPT-4o) to ~83% (o1) to near-saturation; GPQA and ARC-AGI fell in sequence. Benchmarks now specify 'with thinking' the way chip specs specify precision.
What it did to infrastructure
A hard question can burn tens of thousands of hidden tokens — decode-heavy, KV-cache-hungry work that lands exactly on inference's expensive phase. The 2025 hardware wave reads as a direct response: 288 GB GPUs (B300), inference-first TPUs (Ironwood), disaggregated serving stacks. When executives say demand 'went vertical' in 2025, reasoning tokens are a large share of what they are metering.
Open questions worth tracking
Does the approach generalize beyond checkable domains (law? strategy?) — early agentic results say partially. Do visible chains of thought reflect real computation (interpretability debates continue)? And the economic one: users pay per answer, labs pay per token thought — pricing thinking is the industry's newest margin problem. Watch it on our Trends and News pages.
