Reasoning models think before they speak: why that changes everything
In March 2025, Anthropic released Claude 3.7 Sonnet with an optional "extended thinking" mode. OpenAI had introduced its o1 model series in late 2024 and was refining it with o3 and o4-mini. The trend was clear: a new class of language model was arriving, one that does not produce an answer immediately. Instead, it spends time working through a problem in a scratchpad before committing to a response. This sounds like a minor implementation detail. It is not.
What standard models actually do
A standard LLM samples one token at a time from a conditional distribution:
No scratchpad, no revision step — just a single autoregressive chain. Fast and cheap, but hard problems need more than one forward pass.
The problem is that generating the correct answer to a hard problem in a single pass requires that the answer be directly recoverable from the training distribution. For tasks like writing emails, summarising documents, or translating text, that works well. For tasks involving multi-step logical deduction, mathematical proof, or software architecture with competing constraints, a single pass often produces plausible-sounding but incorrect results.
What reasoning models do differently
Reasoning models generate a chain of thought before producing their final answer. In practice this looks like the model writing out intermediate steps, checking conditions, revising earlier conclusions, and arriving at an answer through a process that is visible in the output. The key architectural insight is that the computation budget is flexible. Hard problems get more thinking steps. Simple problems get fewer.
Claude 3.7's extended thinking mode takes this further by allowing users to set a token budget for the thinking process explicitly. You can tell the model to think longer on a problem you know is hard, or use standard mode for tasks where extra reasoning time does not help. The thinking tokens are generated but not shown to the user by default; they are an internal workspace.
The core insight is that intelligence is not just about knowing things. It is about being able to allocate effort proportionally to problem difficulty. That is what reasoning models add.
Performance on hard tasks
The gains are most pronounced on tasks that require multi-step reasoning. AIME (high school mathematics olympiad problems), competitive programming, and graduate-level science questions all show dramatic improvement compared to standard models. The gain is not marginal.
The cost tradeoff
Extended thinking is not free. If standard mode costs per query ( = price per token), reasoning mode adds hidden thinking tokens:
For most applications this does not matter. If you are building a tool that helps an engineer debug a complex distributed systems failure once a week, the extra API cost is irrelevant compared to the value of getting the right answer. If you are processing millions of customer support tickets per day, you would not use extended thinking for routine queries.
Is this actually reasoning?
The philosophical question is worth engaging with honestly. These models are still fundamentally pattern-matching systems trained on text. The "thinking" that happens in the extended thinking mode is also token prediction, just in a longer sequence. There is no guarantee the intermediate steps reflect genuine logical derivation rather than learned patterns that superficially resemble reasoning.
What is unambiguous is that the outputs are more often correct on hard problems. Whether you call that reasoning, systematic search, or learned heuristics is partly a semantics question. The practical value is not. For developers building systems where correctness matters more than latency, reasoning models represent a genuine step up in reliability.