Reasoning Tokens Explained — How o1/o3/DeepSeek-R1 "Think"

What reasoning tokens are

Reasoning models generate an internal chain of thought before producing their final answer. OpenAI calls this "reasoning tokens" (shown as "thinking" in the API response). DeepSeek R1 outputs its <think>...</think> block. These reasoning tokens: cost money (output pricing), are not visible to users by default, and substantially improve quality on hard problems.

Cost impact

For a hard math problem, o3 might generate 2,000 reasoning tokens and 100 answer tokens. You pay output price for all 2,100 tokens. On simpler questions, o3 generates fewer reasoning tokens (50–200). Some APIs let you cap reasoning token budget — lower cap = less thinking = lower cost but possibly worse quality.

When reasoning models are worth the premium

Reasoning models excel at: mathematical proof verification, competitive programming, multi-step logical deduction, long-horizon planning. They're often WORSE than standard models at: simple Q&A, summarization, creative writing, tasks where overthinking hurts. Don't default to reasoning models for everything.

DeepSeek R1 — open reasoning alternative

DeepSeek R1 is an open-weight reasoning model (MIT license) with performance comparable to o3 on math/code benchmarks. Its reasoning tokens are visible in the response. Via API (Together, Fireworks), it costs significantly less than o3. Self-hostable for teams with GPU infrastructure.