AI summaries of long articles: what to keep and what to check

An AI summary of a long article is useful only if every number stays the one in the original: the model compresses the text, not the facts. The rest of the work is review.

The prompt that keeps the numbers

Ask for three things: the numbers with the sentence that explains them, the list of what the summary drops, and a ban on adding facts. Without the second item the summary looks complete and is not.

• Numbers with their sentence: every figure carries its subject and unit of measure.
• Declared cuts: the model lists the sections and details it left out, so the reader knows what is missing.
• No inference: when a fact is absent the answer is "not in the text", not an estimate.

import { generateText } from 'ai';

const summary = await generateText({
  model: 'anthropic/claude-haiku-5.5',
  system:
    'Summarise the article. Keep every number with the sentence that explains it. ' +
    'List what you left out. Never add facts that are not in the text.',
  prompt: article,
  maxOutputTokens: 700,
});

console.log(summary.text);
console.log(summary.usage);

Caching changes the bill

An article read by a hundred people is a hundred requests with the same prefix: a cache read costs a tenth of the input price and pays for itself after a single reuse, while the write costs 1.25 times. The saving exists only when the prefix does not change between requests.

Prices read on 1 November 2026 from both providers' pages; they change without notice, so the date is part of the figure.

Where a summary goes wrong

• Negations: "the plan does not cover refunds" becomes "covers refunds" when the model compresses.
• Orphan numbers: a figure without its subject stays plausible and false.
• Quotes: quotation marks survive the rewrite and attribute sentences nobody said.

The limitation

A summary cannot be scored without a reference: this site has no set of hand-approved summaries, so we publish no accuracy figure. The tokenizer matters too: the same text costs more tokens in Italian than in English, and the 30% Anthropic states applies to models from 4.7 onwards.

Where to continue

The cost of long context is in long context: real costs, streaming the answer in streaming with the AI SDK, comparing models in evaluating a model for code. Records worth opening: 111,985 components, 1,556 blocks and the 21st.dev sidebar.