Skip to the page
Tecyfy

Mastering Chain-of-Draft: Optimizing Amazon Bedrock Efficiency

Chain of Draft asks a model to reason in notes of 5 words or fewer. What the research found, where it fails, and how to test it on Amazon Bedrock.

Tecyfy · Published · Updated · 7 min read

Written in December 2025, when this was news. Models and prices have moved on since: today's are on Model prices.

Chain-of-thoughtChain of ThoughtHaving a model write out the steps of its reasoning before it gives an answer. Once a prompting trick, now built into "thinking" models. prompting makes a model reason step by step before it answers, which raises accuracy on maths and logic but makes every answer long, slow and costly. Chain of Draft is a variation that asks the model to keep each reasoning step to a short draft of 5 words at most. In the paper that introduced it, it kept most of chain-of-thought's accuracy with a fraction of the tokens. This article explains the method, what the numbers show and where it fails, with a version for Amazon Bedrock you can run.

From chain of thought to chain of draft

A chain-of-thought answer reads like a worked solution written out in full:

Jason had 20 lollipops. He gave some to Denny and now has 12 left. So he gave away 8.

A chain-of-draft answer keeps only what a person would jot on paper:

20 − 12 = 8. #### 8

The instruction that does this, from the paper, is short:

Think step by step, but only keep a minimum draft for each thinking step, with 5 words at most.
Return the answer at the end of the response after a separator ####.

Fewer output tokensTokenThe small piece of text a language model reads and writes, often part of a word. Prices, limits and the context window are all counted in tokens. means a lower token cost, because there is less to generate, and a faster answer, because a model writes its answer one token at a time.

What the research found

The method comes from a February 2025 paper by researchers at Zoom, Chain of Draft: Thinking Faster by Writing Less. On GSM8K, a set of school maths word problems, it tested two leading models of the time, GPT-4o and Claude 3.5 Sonnet:

PromptAccuracyOutput tokensLatency
Answer directly53–65%about 1under 1 s
Chain of thoughtabout 95%about 2003–4 s
Chain of draftabout 91%about 401–1.6 s

So chain of draft gave up about 4 points of accuracy for about 80% fewer tokens. On some commonsense tasks it matched or beat chain of thought; at best it used 7.6% of the tokens.

Where it fails

The same paper is clear about the limits:

  • Without examples it falls apart. The results above used few-shotFew-shot LearningPutting a few worked examples of the task in the prompt, so the model copies their format, labels and tone without any training. examples that showed the drafting style. With the instruction alone, Claude 3.5 Sonnet's accuracy on GSM8K fell to 65.5% against 90.4% for chain of thought. Always include 2 or 3 worked examples in the draft style.
  • Small models do worse. On models under 3 billion parameters the gap to chain of thought was much wider.
  • You lose the explanation. A draft is hard for a person to audit. Where you need a readable reasoning trail, as in regulated work, keep chain of thought.

Chain of draft shortens the reasoning you ask a model to write in its answer; it is a prompt technique, so test it against the model's own settings rather than assuming it helps.

Trying it on Amazon Bedrock

AWS tested chain of draft on Bedrock with a logic puzzle and 2 models. Chain of draft cut the total count by 39% and 76% compared with chain of thought, though the AWS code counted words rather than real tokens, and latency by 52% and 79%. The gain depends heavily on the model, so measure your own.

This version calls the Bedrock Converse API with a few-shot example and reads the real token counts and latency that Bedrock returns, rather than counting words. Choose a model you have access to in the Bedrock console and put its id in MODEL_ID.

import boto3

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")
MODEL_ID = "your-model-id"  # from the Bedrock console

SYSTEM = (
    "Think step by step, but only keep a minimum draft for each thinking step, "
    "with 5 words at most. Return the answer at the end after a separator ####."
)

# One worked example teaches the drafting style; the paper found it matters.
EXAMPLE_Q = "Jason had 20 lollipops. He gave Denny some. Now he has 12. How many did he give?"
EXAMPLE_A = "20 - 12 = 8. #### 8"

QUESTION = (
    "Three boxes are labelled 'Red', 'Blue' and 'Red and Blue'. Every label is wrong. "
    "You take one ball from the box labelled 'Red and Blue' and it is red. "
    "What does each box contain?"
)

response = bedrock.converse(
    modelId=MODEL_ID,
    system=[{"text": SYSTEM}],
    messages=[
        {"role": "user", "content": [{"text": EXAMPLE_Q}]},
        {"role": "assistant", "content": [{"text": EXAMPLE_A}]},
        {"role": "user", "content": [{"text": QUESTION}]},
    ],
    inferenceConfig={"maxTokens": 300},
)

print(response["output"]["message"]["content"][0]["text"])
usage = response["usage"]
print(f"Output tokens: {usage['outputTokens']}, latency: {response['metrics']['latencyMs']} ms")

A good draft for the puzzle looks like this:

'Red and Blue' label wrong.
Ball red, so box: red.
'Blue' box: not blue, not red-only.
'Blue' box: red and blue.
'Red' box: blue.
#### 'Red and Blue' box holds red; 'Blue' box holds red and blue; 'Red' box holds blue.

To compare fairly, run the same questions with a chain-of-thought system prompt ("Think step by step, then give the answer after ####") and with no reasoning at all, and compare accuracy as well as outputTokens and latencyMs. A shorter answer is only a saving if it is still right.

Tecyfy Takeaway

Chain of draft asks a model to reason in short notes instead of full sentences. With a few worked examples it keeps most of chain-of-thought's accuracy for a fraction of the tokens, which makes it worth testing wherever reasoning prompts drive your costs or your response times. Without examples, on small models, or where a person must read the reasoning, stay with chain of thought.

Questions

What is chain of draft prompting?
Chain of draft asks a model to reason step by step but keep each step to a short draft of 5 words at most, with the answer after a #### separator. It comes from a February 2025 paper by researchers at Zoom.
Is chain of draft as accurate as chain of thought?
Nearly, with examples. On the GSM8K maths problems it scored about 91% against about 95% for chain of thought, while using about 40 output tokens instead of about 200, so about 80% fewer.
Why does chain of draft need few-shot examples?
Without worked examples in the draft style, accuracy falls sharply: Claude 3.5 Sonnet dropped to 65.5% on GSM8K against 90.4% for chain of thought. Include 2 or 3 examples written as short drafts.
When should I not use chain of draft?
On small models under about 3 billion parameters, where the gap to chain of thought was much wider, and wherever a person must read and audit the reasoning, as in regulated work.
How do I test chain of draft on Amazon Bedrock?
Call the Converse API with the chain-of-draft system prompt and one worked example, then compare outputTokens, latencyMs and accuracy against a chain-of-thought prompt on the same questions. AWS measured 39% and 76% lower totals on 2 models, counted in words rather than tokens, so the gain depends on the model.