03Blog
From the lab
Explainers and practical guides on how AI systems work, and how to test them before you rely on them.
Explainer · 26 Sep 2026
How reasoning models decide how long to think
Reasoning models can think for a few words or for thousands. How they decide, why it matters for cost and accuracy, and where the approach breaks.
Read it- How to test an AI assistant before you launch itA practical method for small teams: realistic test cases, the right judges, and measures that catch the failures that matter. No special tooling required.3 min
- What calibration means, and why it mattersA calibrated model is right about as often as it claims to be. Why that matters more than raw accuracy for anything people rely on, and how to check it.2 min
- Why the same question gets different answersAsk a model the same thing twice, or in slightly different words, and the answer can change. The four main causes, and what to do about each.2 min
- How AI models change after launchA deployed AI system keeps changing through vendor updates, prompt edits, new data sources and fine-tuning. How to notice when behaviour shifts, and what to set up before it does.2 min
- What an AI model really costs to runPrice per token is only the starting point. How to work out the cost of a correct answer, including thinking tokens, retries and the people who handle what the model can't.2 min