Quality is not a free property of AI-assisted development. Teams that cut API costs without tracking output quality can discover the damage only after customers complain, creating a delayed and difficult-to-locate regression. section 9.0
Treat quality as a workflow-dependent tradeoff against latency and cost. A single middle-tier configuration often performs poorly across interactive coding, batch processing, complex reasoning, and lightweight completion. section 9.5.4
Evaluate quality against the task rather than a single global benchmark:
- Use stronger models for complex reasoning, multi-file generation, deep CI review, and overnight test generation when quality matters more than latency. section 9.5.4
- Accept lower quality for simple, short tasks such as tab completion, classification, boilerplate, and immediate first-pass CI review. section 9.5.4 section 9.3.5
- Test quantized models on representative work: degradation may be negligible for short completions but visible in multi-line functions and test suites. section 9.3.5
Choosing a local coding model
A team comparing Q4_K_M with Q6_K found less than a 2% acceptance-rate difference on tab completion, inline edits, and test scaffolding, while Q4_K_M preserved more RAM for the IDE and containers. For function-length generation, it stayed at Q4_K_M or above because lower levels produced inconsistent names, subtle control-flow bugs, and forgotten constraints. section 9.3.5
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
|
Use the workflow matrix to match quality requirements with latency and cost instead of applying one model configuration everywhere. section 9.5.4









