Open dataset of what AI coding agents cost: token pricing, per-million rates, and every published figure (prices, percentages, multiples, token counts and durations) as JSON and CSV. Figures are one row each, carrying the sentence they came from. Plus the field notes behind the numbers.
View the Project on GitHub xyzs996/ai-coding-field-notes
Code Review
3 of the 30 write-ups here are tagged Code Review. Every figure quoted below is in the figures table with the sentence it came from.
- 80% — When the Claude Code team decided to slash 80% of their system prompts, most developers expected the model to lose its edge in complex engineering tasks. →
- 80% — Stripping away that redundant 80% removes the cognitive drag holding the model back, freeing native reasoning capacity and cutting the token burn. →
- $1.43 — For instance, GPT-5.6 Sol, while more expensive at $1.43 per run, shows superior performance with a 43.1% accuracy rate in the same ReactBench tests, which shows that cheaper models may save money upfront but can lead to longer development cycles due to frequent errors and rework. →
- $1.43 — In contrast, GPT-5.6 Sol, at $1.43 per run, achieves 43.1% accuracy in the same tests, suggesting that while cheaper models may save money upfront, they often result in longer, more costly development processes. →
- 95% — That 95% export-issue figure gets quoted everywhere. →
- 60% — It integrates with GitHub Actions and CircleCI, and its structured review comments cut false positives by 60% — which is the number I’d actually optimize for, because a review tool that cries wolf gets muted inside a week, and a muted tool reviews nothing at all while still showing green on every dashboard I own. →
All figures, 306 rows
The write-ups
Open Code Review is an open-source review tool built for AI-assisted development, and in benchmark tests spanning 200 real pull requests across 50 open-source repositories it scored higher on both…
Fable 5, the cheapest option at $9.05 per run, delivers only 41.2% accuracy in React projects.
When the Claude Code team decided to slash 80% of their system prompts, most developers expected the model to lose its edge in complex engineering tasks.
All 30 write-ups