I've seen teams burn six-figure cloud bills on "optimized" AI coding setups that deliver negative ROI. The pattern is always the same: cargo cult adoption of "best practices" without measuring actual cost versus developer velocity.
Most common offenders:
* **Massive context windows as default**: Feeding entire codebases into GPT-4 Turbo 128K for every minor refactor. At $0.06 per 1K output tokens, a few engineers can easily hit $500/day in API costs. If you're not using cheaper models for 80% of tasks, you're just lazy.
* **Over-engineered custom instructions**: I reviewed a team's 5000-character "perfect prompt" that required maintaining a separate documentation repo. Their monthly Claude Team subscription cost more than their actual productivity gain. Simpler instructions with targeted examples almost always outperform verbose manifestos.
* **Automated code review pipelines that run on every PR**: One startup spent $2200/month on GitHub Copilot Business plus GPT-4 review bots. The bots generated more noise than actionable feedback, increasing review cycle time by 40%.
The only metrics that matter:
- Cost per accepted change (API calls + subscriptions / merged PRs)
- Time from prompt to production-ready code
- Regression rate from AI-assisted versus human-only work
If you're not tracking these, you're optimizing for conference talk points, not business value. Start with the cheapest model that works, measure rigorously, and only scale complexity when the numbers justify it.
Example of a cost-aware setup for AWS workloads:
```yaml
# assistant_strategy.yaml
default_model: claude-3-haiku # $0.25 per 1M input tokens
escalation_model: claude-3-sonnet # only for complex architecture
context_limits:
file_review: 5_files_max
total_tokens: 8000
cost_triggers:
daily_alert_threshold: $50
require_approval_above: $200
```
This cut one team's monthly AI spend from $3,400 to $780 while maintaining 95% of their velocity gains.
cost optimization, not cost cutting