Reasoning scores keep climbing while per‑token compute scarfingly drops. GLM‑5.2 hits 99.2% on AIME 2026 using about 40 billion active parameters; Qwen3.5 scores 91.3% with 17 billion active; Deep Seek V4‑Flash runs 13 billion active per token. For scale, GPT‑4 was rumored at ~280 billion active parameters in 2023 and barely solved an AIME problem.

When you shift from math and code to factual recall, the picture flips. On Simple QA, the leader is Gemini 2.5 Pro at 53%, so even money‑buyable recall still misses half the questions. Qwen3.5 4 B and 9 B register hallucination rates of 80–82% on a knowledge benchmark, meaning most of the time they fabricate facts. Ask for a minor mathematician’s birth year and you get a confident, plausible wrong answer.

The trade is deliberate: knowledge takes space. Research shows ~two bits of factual knowledge per parameter. If you want a model that knows every Wikipedia birth year, you pay for that in weights, which drives frontier models to trillions. Reasoning compresses better because it’s a small set of procedures applied repeatedly. Phi‑4 (14 B) trained on synthetic textbook data is good at math but bad at trivia, reflecting its training mix.

The solution is a harness: retrieval, tool calls, and web search supply knowledge at runtime, letting the model focus on reasoning. Models that are largely procedural age far slower – facts rot, procedures don’t. A frontier model on a single consumer GPU could deliver frontier‑quality reasoning while leaving knowledge to a lightweight, up‑to‑date database.