ℹ️ CodeLens has shut down. Read more about why we shut down or check out FlouState for developer productivity tracking.
Insights, case studies, and analysis of AI model performance on real-world coding challenges
Testing Anthropic's newly released Claude Haiku 4.5 revealed a surprising paradox: it produced 13,666 tokens (most of all 8 models) but scored 74.4/100, while Sonnet 4.5 wrote 8,425 tokens and scored 89.0. This is what over-engineering looks like in AI-generated code.
We tracked carbon emissions across 173 AI code generations from 28 real developer tasks (5 days post-launch). Surprising finding: same category of work, 90x different carbon footprint. Gemini averages 0.65g CO2 per query, GPT-5 averages 2.31g (3.6x worse). Here's why grid renewables matter more than model efficiency.
More case studies and insights coming soon!
Have a suggestion for a topic? Let us know