43. Model Selection for Analysis

The lesson

Long document analysis (legal, research papers) requires models with 1M+ context windows like Claude Opus 4.8 or Gemini.

Structured reasoning and comparison frameworks benefit from models with explicit reasoning capabilities.

Reasoning models (o3-mini, DeepSeek-R1) show their thinking process, helping verify that conclusions follow from evidence.

For multi-step analysis tasks, prefer models that maintain coherence across long chains of reasoning.

When accuracy is critical, use reasoning models and ask them to show their work for verification.

A legal analyst uses a long-context model to review an entire case bundle, then a reasoning model for comparing conflicting evidence and asking it to show its working.

A researcher at a laboratory institute selects a long-context model for reading whole papers and a reasoning model for statistical comparisons.

A financial analyst picks a long-context model for annual reports and a reasoning model for multi-step ratio analysis where each step must be checkable.

An underwriter chooses a reasoning model for assessing complex risk combinations, requiring the model to show how it reached each conclusion.

A policy researcher at a government department uses a long-context model to review consultation submissions in full, then a reasoning model to compare the arguments.

A clinical audit officer selects a long-context model for reviewing full guidelines and a reasoning model for tracing where practice diverged from them.

A resource analyst at a mining company picks a reasoning model for multi-step grade estimates, asking it to display the calculation path for checking.

An investigative reporter uses a long-context model to read a large document set, then a reasoning model to compare timelines across the sources.

An environmental consultant selects a reasoning model for multi-step impact comparisons, requiring each conclusion to show the evidence behind it.

An assessment analyst at an education authority uses a reasoning model to compare moderation results across cohorts, asking it to show the working for verification.

Check yourself