Models that think before they answer

For years, better models meant bigger models trained on more data. Around 2024 a second lever became clear: let the model spend more computation at the moment of answering. Instead of producing the answer directly, it produces a long internal chain of steps, checks itself, explores alternatives and then answers. Performance on hard problems improves substantially, and it improves predictably with the amount of thinking allowed.

Why it works at all

A language model produces one token at a time, and each token gets a fixed amount of computation. A hard problem that needs many logical steps cannot be compressed into that fixed budget. Writing the steps out gives the model more computation and, crucially, a place to put intermediate results, which it can then read back. The chain of thought is scratch paper.

The second ingredient is training on whether the final answer was right, rather than on imitating human text. When a problem has a checkable answer, a model can be trained by rewarding chains that reach it. That is why the gains are concentrated in mathematics, code and formal reasoning, where correctness is machine-checkable, and thinner in domains where it is not.

What this does to cost and latency

A reasoning model can generate thousands of hidden tokens before its first visible word. You pay for all of them, and you wait for all of them. A question answered in two seconds by a standard model may take forty and cost twenty times as much.

That changes product design. The right pattern is routing: easy questions to a fast model, hard ones to a reasoning model, with the decision made by a cheap classifier or by the user. Applications that send everything to the expensive path burn their budget on questions that did not need it.

The limits people keep rediscovering

Longer thinking does not fix missing knowledge. If the model does not know a fact, more steps produce a more elaborate wrong answer. Retrieval fixes this; reasoning does not.

The chain is not a transcript of the reasoning. The visible steps are text the model generated, and they can be plausible and disconnected from whatever computation produced the answer. Treating the chain as an explanation is a mistake with real consequences in regulated settings.

Performance degrades on genuinely novel structure. Benchmarks improve fast, partly because problem types leak into training data. The gap between benchmark scores and performance on a problem nobody has posted online is real and hard to measure.

The efficiency question

The DeepSeek release in early 2025 showed that a competitive reasoning model could be trained for far less than the assumed cost. The techniques, distillation from a stronger model, mixture-of-experts architectures that activate only part of the network per token, and careful reinforcement learning on verifiable tasks, are now standard. The practical consequence is that capability per euro has fallen fast, and continues to.

This has a straightforward implication for anyone building on models: assume the price of a given capability level falls significantly each year, and do not architect around today's cost of the frontier.

What to watch

Watch whether reasoning gains transfer to domains without checkable answers, which would be the genuinely surprising result. Watch the cost of a fixed benchmark score over time, which is the cleanest efficiency measure. And watch for regulated industries requiring faithful explanations, because that is a requirement no current model architecture meets.