Real World Example

Next-token prediction is directly visible in GitHub Copilot code completion, Gmail Smart Compose and Notion AI "continue writing". The editor sends the text before the cursor; the model continues it. Here is a ShopEase admin feature that continues a half-written product description:

import anthropic

client = anthropic.Anthropic()

draft = "Handwoven cotton kurta in indigo blue. Breathable fabric, ideal for"

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=120,
    system="You continue product descriptions for an Indian e-commerce store. "
           "Return only the continuation (1-2 sentences), no quotes, no preamble.",
    messages=[{"role": "user", "content": f"Continue this text:\n{draft}"}],
)
print(draft, response.content[0].text)
print("stop_reason:", response.stop_reason, "| output tokens:", response.usage.output_tokens)

python continue_description.py

Handwoven cotton kurta in indigo blue. Breathable fabric, ideal for summer office days and festive evenings alike. Pair it with white churidar or jeans for an easy, elegant look.
stop_reason: end_turn | output tokens: 31

• The model writes fluent text but knows nothing about the actual kurta: a seller must review claims like "breathable" (hallucination risk).
• stop_reason is end_turn: the model decided it was done. If it were max_tokens the text was cut off.