AI Reasoning Models & Test-Time Compute
Standard models answer immediately with a fixed amount of compute per token. Reasoning models spend additional “thinking” tokens working through a problem before they answer.
From pre-training scale to test-time compute
For years, capability grew mainly by pre-training: more parameters, more data, larger clusters. Reasoning models add a second axis, inference-time compute. By training models to work through intermediate steps, check their own work, and abandon unpromising paths, providers get better results on problems that need several dependent steps.
It is now a setting, not a separate model
Earlier generations shipped dedicated reasoning models alongside standard chat models. Today the major providers fold this into their flagship tiers. Anthropic’s documentation describes adaptive thinking, where the model decides how much to think and an effort parameter steers it. The practical question is no longer which reasoning model to pick, but how much reasoning effort a task deserves.
When to turn reasoning up
- High-stakes logic: checking proofs, contract logic, or safety-critical rules.
- Hard quantitative problems: physics, chemistry, and competition-style math.
- Root-cause analysis: diagnosing bugs that span several systems.
For brainstorming, quick lookups, or short drafts, higher reasoning effort mostly adds latency and cost. Use the lowest effort that solves your task.