JMN Field Manual

02.4 — Building Agents

Model Selection Guide

Every model option produces deterministic answers — temperature is fixed at 0 across the board, so picking a model is purely a question of capability, output length, and cost, never "creativity."

6.1
The model menu grouped by Claude and ChatGPT

FIG. 6.1 — Nine models across two providers, grouped in the same menu.

ModelProviderMax output / runPick it for
Opus 4.7Claude128,000 tokensThe hardest reasoning or longest single answer in the lineup
Sonnet 4.6Claude64,000 tokensThe default — strong general-purpose reading and extraction
Haiku 4.5Claude64,000 tokensHigh-volume, simple columns where speed and cost matter more than depth
GPT-5.2OpenAI128,000 tokensA second opinion from outside the Claude family, or client preference
GPT-5 miniOpenAI128,000 tokensLighter GPT-5-family runs at lower cost
GPT-5 nanoOpenAI128,000 tokensThe cheapest, fastest GPT-5-family option
GPT-4.1OpenAI32,000 tokensPrior-generation GPT, shorter answers
GPT-4.1 miniOpenAI32,000 tokensLighter prior-generation GPT runs
GPT-4.1 nanoOpenAI32,000 tokensCheapest prior-generation GPT option

Note

"Max output" is a hard ceiling on how long that model's answer can be, not a target — most cells return far fewer tokens. The platform sets a safe default per property and lets it be raised up to this ceiling when a genuinely long answer (a full summary, a generated table) needs the room.

Tip

Default to Sonnet 4.6 until you've actually seen it get something wrong on your document type. Reach for Opus 4.7 when the document is dense or the question requires real reasoning across it — and remember Full Doc context mode (see Tools & How They Work) is a Claude-only option, so it's not available if you switch a column to a ChatGPT model.