Side-by-Side Comparison
| Chain-of-Thought | Chain of Draft | |
|---|---|---|
| Instruction | "Think step by step…" | "Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most…" |
| Step format | Full sentences and explanations | Numbers, equations, keywords |
| GSM8K, GPT-4o | 95.4% · ~205 tokens · ~4.2 s | 91.1% · ~44 tokens · ~1.0 s |
| GSM8K, Claude 3.5 Sonnet | 95.8% · ~190 tokens · ~3.1 s | 91.4% · ~40 tokens · ~1.6 s |
| Zero-shot | Works well | Weak; needs draft-style examples |
| Small models | Helps | Larger gap to CoT under ~3B parameters |
| Readability | Easy for humans to audit | Terse; harder to review |
| Introduced by | Wei et al., 2022 | Xu et al., 2025 |
Benchmark numbers are from the Chain of Draft paper (Xu et al., 2025) and will differ on newer models and your own tasks.
The Same Problem, Two Ways
Chain-of-Thought
Q: Jason had 20 lollipops. He gave Denny some. Now he has 12. How many did he give Denny? A: Jason started with 20 lollipops. After giving some to Denny he has 12 left. The number he gave away is the difference between what he started with and what he has now: 20 - 12 = 8. So he gave Denny 8 lollipops. #### 8
Chain of Draft
Q: Jason had 20 lollipops. He gave Denny some. Now he has 12. How many did he give Denny? A: 20 - x = 12; x = 8. #### 8
Decision Guide
- Paying per token at scale, or users waiting on responses? Chain of Draft, with 2–3 draft-style examples.
- Need an explanation a person will read or audit? Chain-of-Thought.
- Accuracy is critical? Chain-of-Thought, or Chain of Draft plus Self-Consistency, then compare on a test set.
- Small or local model? Chain-of-Thought.
- Reasoning model? Skip explicit chains; state the problem clearly and ask for a concise answer.
Frequently Asked Questions
What is the difference between Chain-of-Thought and Chain of Draft?
Both ask the model to reason step by step before answering. Chain-of-Thought lets each step be a full explanation; Chain of Draft limits each step to a minimal draft of about five words. The result is the same structure with far fewer tokens.
How much cheaper is Chain of Draft?
In the Chain of Draft paper, reasoning on the GSM8K math benchmark dropped from about 205 to 44 tokens per answer on GPT-4o and from about 190 to 40 on Claude 3.5 Sonnet, roughly an 80% reduction, with latency falling by about 60–75%.
Is Chain of Draft less accurate?
Sometimes slightly. On GSM8K accuracy fell from about 95% to about 91%. On date understanding and sports understanding tasks, Chain of Draft matched or beat Chain-of-Thought. Test both on your own tasks before switching.
Which should I use with reasoning models?
Neither needs much emphasis. Reasoning models already think internally before answering, so an explicit visible chain adds little. If you do want visible reasoning from them, Chain of Draft keeps the output short.
Can I combine Chain of Draft with Self-Consistency?
Yes, and it is a good fit. Self-Consistency samples several reasoning paths and takes a majority vote, which multiplies cost. Short drafts make sampling five or more paths affordable.