Crabhaus

What it is

Buying answer quality at inference time by letting a model generate and check far more tokens before replying. Reasoning models - OpenAI's o-series, DeepSeek-R1, Claude's extended thinking - are trained so accuracy climbs as they 'think' longer, making per-query token spend a scaling axis alongside bigger pretraining runs.

How it works

RL on problems with checkable answers (math, code) teaches the model to emit a long private chain of thought: try an approach, catch contradictions, backtrack, verify. DeepSeek-R1 showed this emerges from simple rule-based correctness rewards. Compute can go sequential (longer chains) or parallel (sample many answers, pick by vote or verifier); providers expose it as a reasoning-effort knob.

A concrete example

Give a model a competition math problem on a tight budget and it pattern-matches to a plausible wrong answer. The same weights with a 10,000-token thinking budget work the problem - an inconsistency caught, arithmetic re-verified, correct answer returned. Nothing changed but the compute the query bought: ~50x the tokens, cost, and latency.

Why it matters

It rewrites inference economics: cost and latency now depend on how hard the model thinks, not just answer length - so route easy calls to fast models and spend thinking budgets on hard ones. Snell et al. showed small models with generous test-time compute beating ~14x larger ones, shifting value from parameter count to inference infrastructure.