There is no best model. There is a best model for this task, at this length, at this price, and the gap between choosing well and choosing by habit is enormous.
Match the model to the shape of the task
- Small edits, renames, copy changes: the fastest capable model. Reasoning here is money set on fire.
- Architecture and multi-file refactors: a deep reasoning model, once, with the full picture — cheaper than three fast attempts.
- Design work: a strong multimodal model that can look at the reference and the result.
- Long transcripts: whatever holds the context without degrading, even if it is slower.
- Bulk extraction and classification: small, cheap, structured output.
Cost intuition
Reasoning tokens are output tokens. A model that thinks for thirty seconds before a two-line edit can cost twenty times a direct answer that would have been identical. Save the deliberation for decisions that are actually hard to reverse.
Context is a budget, not a bucket
Filling a million-token window because you can is how quality quietly drops. Send the goal, the relevant files, and the last few turns. Everything else is noise you are paying to have ignored.
Still wondering something?
The assistant answers from these pages only.