Content summarization

When to use this scenario

Long-form summarization covers research papers, earnings call transcripts, legal filings, newsletter digests, and meeting recordings converted to text. The defining challenge is context window: a 120-page annual report exceeds 100K tokens. Models that truncate silently produce coherent-sounding but factually incomplete summaries — a risk that compounds when humans stop reading the source.

Gemini 2.5 Pro's 2M-context window handles the longest documents without chunking. At $4/million input tokens, summarizing 50,000 documents per month (avg 8K tokens each) runs roughly $1,600 in input cost. Claude Sonnet is a credible fallback at half the price with 200K context — sufficient for most individual documents but inadequate for book-length inputs.

Output is short (400 tokens avg), so output cost rarely dominates. The real cost driver is input length. Benchmark your P95 document size before committing to a model.

Common pitfalls

• Chunking long documents and summarizing chunks independently — key conclusions often span section boundaries and get dropped
• Using output token count as a proxy for quality; summaries that omit critical caveats are shorter and score higher on ROUGE but fail real tasks
• Ignoring cached-prefix discounts: if many documents share a long system prompt, providers with prompt caching (Anthropic, Google) can cut input cost 60–90%
• Assuming all models handle 8K inputs equally — some providers still count padding tokens against rate limits