Reasoning models
The models that think before they answer
What are reasoning models
Reasoning models generate internal "thinking" tokens before producing their final answer. OpenAI's o1 (September 2024) was the first major commercial reasoning model. DeepSeek R1 (January 2025) proved the approach works with open weights.
The key difference from standard models: a reasoning model spends compute on a chain-of-thought process that the user may or may not see. This thinking step improves accuracy on complex tasks — mathematics, logic, multi-step problems — but increases both latency and cost, because thinking tokens are billed at output rates.
Why it matters
Reasoning models changed the price-performance landscape. o1 costs $15/1M input — five times more than GPT-4o — but solves problems GPT-4o cannot. The trade-off is explicit: more thinking, more tokens, more cost, better answers. sourc.dev tracks extended thinking support as a capability flag and includes reasoning models in the SVI computation.
Verified March 2026 · Source: OpenAI o1 announcement, DeepSeek R1 paper