The latency-cost-quality triangle frames AI-assisted development as a three-way trade-off: low latency, low cost, and high quality cannot be maximized simultaneously. Each workflow should occupy its own position rather than inherit one universal configuration. section 9.5.4
Use the triangle's edges to match configuration to workflow:
- Choose low latency plus high quality for interactive pair-programming; expect higher expense from fast, full-context agent sessions. section 9.5.4
- Choose low cost plus high quality for background refactoring, deep CI review, and overnight test generation; accept asynchronous delivery. section 9.5.4
- Choose low cost plus low latency for tab completion, inline edits, and boilerplate; accept lower performance on complex reasoning. section 9.5.4
Table 9.7 Matching model to the usecase
|
Technique
|
Properties
|
Model
|
|
Interactive pair-programming
|
Latency + Quality
|
Sonnet, agent mode
|
|
Tab completion
|
Cost + Latency
|
Haiku or local model
|
|
Background refactoring
|
Cost + Quality
|
Sonnet, async/batch
|
|
CI review (first pass)
|
Cost + Latency
|
Haiku, immediate
|
|
CI review (deep pass)
|
Cost + Quality
|
Sonnet, batch API
|
|
Test generation
|
Cost + Quality
|
Sonnet, overnight
|
|
Documentation
|
Cost + Latency
|
Haiku, on-demand
|
Compare recommended model configurations across interactive development, CI review, refactoring, testing, and documentation workflows. section 9.5.4
Why one medium configuration fails
A single mid-tier model applied everywhere becomes too slow for tab completion, too expensive for batch CI, and too weak for complex architecture sessions. Per-workflow positioning avoids mediocrity across all three dimensions. section 9.5.4
