Text Classification with Claude API (Python)

Classify text with Claude in Python: single-label, multi-label, zero-shot, and batch classification. Working code examples with structured JSON output and accuracy tips.

Claude's instruction-following ability makes it an excellent zero-shot and few-shot text classifier. This guide covers single-label, multi-label, and batch classification with structured JSON output.

Installation

pip install anthropic

Single-label classification (zero-shot)

import anthropic
import json

client = anthropic.Anthropic()

LABELS = ["positive", "neutral", "negative"]

def classify_sentiment(text: str) -> str:
    system = f"""You are a sentiment classifier.
Return ONLY a JSON object with a single key "label".
The label must be one of: {LABELS}.
No explanation. No markdown. Raw JSON only."""

    response = client.messages.create(
        model="claude-haiku-4-5-20251001",  # fast + cheap for classification
        max_tokens=32,
        temperature=0,  # deterministic
        system=system,
        messages=[{"role": "user", "content": text}]
    )
    result = json.loads(response.content[0].text)
    return result["label"]

print(classify_sentiment("The new API is incredibly easy to use!"))
# → "positive"
print(classify_sentiment("The package arrived yesterday."))
# → "neutral"

Multi-label classification

TOPIC_LABELS = ["billing", "technical", "account", "feature-request", "bug", "general"]

def classify_support_ticket(ticket: str) -> list[str]:
    system = f"""You are a support ticket classifier.
Return ONLY a JSON object with key "labels" containing an array of applicable labels.
Labels must be from this list: {TOPIC_LABELS}.
Return an empty array if no category fits.
No explanation. Raw JSON only."""

    response = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=64,
        temperature=0,
        system=system,
        messages=[{"role": "user", "content": ticket}]
    )
    result = json.loads(response.content[0].text)
    # Validate against allowed labels
    return [l for l in result["labels"] if l in TOPIC_LABELS]

ticket = "I was charged twice this month and I can't log into my account."
print(classify_support_ticket(ticket))
# → ["billing", "account"]

Few-shot classification (higher accuracy)

def classify_with_examples(text: str, labels: list[str], examples: list[dict]) -> str:
    """
    examples: [{"text": "...", "label": "..."}, ...]
    Provide 2-5 examples per label for best accuracy.
    """
    examples_str = "
".join(
        f'Text: "{ex["text"]}"
Label: {ex["label"]}'
        for ex in examples
    )
    system = f"""You are a text classifier.
Allowed labels: {labels}
Return ONLY a JSON object with key "label".

Examples:
{examples_str}"""

    response = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=32,
        temperature=0,
        system=system,
        messages=[{"role": "user", "content": f'Text: "{text}"
Label:'}]
    )
    result = json.loads(response.content[0].text)
    return result["label"]

examples = [
    {"text": "App crashes on launch", "label": "bug"},
    {"text": "Please add dark mode", "label": "feature-request"},
    {"text": "Can't reset my password", "label": "account"},
]
print(classify_with_examples("Login page freezes after update", ["bug", "feature-request", "account"], examples))
# → "bug"

Batch classification (50% cost via Batch API)

def batch_classify(texts: list[str], labels: list[str]) -> str:
    """Submit up to 100K requests at 50% cost. Returns batch ID to poll."""
    system = f"""You are a text classifier.
Return ONLY a JSON object with key "label".
Allowed labels: {labels}
No explanation. Raw JSON only."""

    requests_list = [
        {
            "custom_id": f"item_{i}",
            "params": {
                "model": "claude-haiku-4-5-20251001",
                "max_tokens": 32,
                "temperature": 0,
                "system": system,
                "messages": [{"role": "user", "content": text}]
            }
        }
        for i, text in enumerate(texts)
    ]
    batch = client.beta.messages.batches.create(requests=requests_list)
    print(f"Batch submitted: {batch.id} — poll until processing_status == 'ended'")
    return batch.id

# batch_classify(my_texts, ["positive", "neutral", "negative"])

Classification approach comparison

For bulk classification jobs, estimate your costs with the Claude API Cost Calculator. For extracting structured fields (not just labels) from text, see the structured output guide. For sentiment + summarization in one call, see the summarization guide.

Frequently asked questions

Free tools

More examples