Hugging Face's logo

Datasets:
JOHNISKING
/
ai-coding-limit-line

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
coding-agents
claude
claude-code
codex
gemini
antigravity
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
ai-coding-limit-line / README.md
JOHNISKING's picture
JOHNISKING
Soften tests-vs-breakage claim
9f9b44a verified
preview code
|
Raw History Blame Contribute Delete
3.66 kB
metadata 
license: cc-by-4.0
pretty_name: >-
  Where Claude Haiku falls behind Opus (163 coding runs, Claude Code, Codex,
  Antigravity)
language:
  - en
tags:
  - coding-agents
  - claude
  - claude-code
  - codex
  - gemini
  - antigravity
  - benchmark
  - llm
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: ai-coding-limit-line.csv

Where Claude Haiku falls behind Opus: 163 coding runs with hidden tests

163 headless runs (plus 11 excluded) of Claude Code (Haiku 5.5, Sonnet 5.5, Opus 5.5), Codex (gpt-6-luna) and Antigravity CLI (Gemini 4 Argon) on six levels of coding tasks, measured 2026-10-23 to 24 on one Windows 11 laptop.

Page, tables and method: https://tokensave.app/data/ai-coding-limit-line · DOI 10.5281/zenodo.23600477 · License: CC BY 4.0

Test cases passed per run on the hard marked bug, by model, with and without running tests

Key findings

  • Levels 2 to 5 (apps from a spec, and an existing 1,032-line app with six symptom-only bugs plus three features): Haiku, Sonnet and Opus got full marks on all 21 runs each, Gemini 18/18, Codex 17/20.
  • Level 6, three real bugs fixed in the marked Markdown parser in Sep to Oct 2026: the easy and medium bugs were solved by almost every run. No run fully fixed the hard one (7-case upstream test), but the models split: Sonnet passed 6 of 7 in 8 of 10 runs, Opus in 6 of 10, Haiku in 1 of 10. The missed cases are code spans containing a backtick, which the bug report did not mention.
  • Letting the agents run the tests (dependencies preinstalled, node, npm test/npm run, read-only git allowed; 5 runs per Claude model) did not change the cases passed. Fewer runs broke other code with tests (2 of 15 without, one Opus run left marked unable to load; 0 of 15 with), but with so few runs that gap could be chance. Codex was also run with tests allowed, but its Windows sandbox blocked esbuild in all 5 runs (excluded).
  • Average API cost per run (Claude Code's own figure): Haiku $0.026, Sonnet $0.19, Opus $0.45.

Files

Prompts, checkers, reference solutions and every run log are in the GitHub folder: https://github.com/johnisking/token-cost-by-language/tree/main/limit-line

  • ai-coding-limit-line.csv / .json: one row per run (score, full marks, seconds, tokens, cost, notes; level 5 and 6 detail columns). excluded marks runs stopped by the harness or by quota.
  • prompts/: every prompt (tier2 to tier5b, tier6_t1 to t3, and tier6_t3_dev for the tests-allowed runs).
  • checkers/: the hidden-test checkers (Playwright, Python) and the tier-6 scorer. ref/, tier5a/, tier5b/: reference solutions, the planted-bug base app and broken variants used to validate the checkers.
  • results/: raw checker output per level.
  • runs/<tool>[dev]_t<level>_r<n>/ (dev = tests allowed): each run's log (run.jsonl or run.json), seconds.txt and the code it produced (work/; for level 6 only work/src). The PC user name is replaced by USER.
  • build_dataset.py: builds the CSV/JSON from the results and logs.

Notes

  • Level 6 scoring: a clean checkout of the commit before the upstream fix plus only the run's src/, the upstream fix's spec file added, then marked's spec suite. 1 point if the hidden spec passes and no other spec test newly fails.
  • Antigravity's headless mode ends a turn at the first denied command. Gemini runs used read-only allow-rules; the level 6 runs were repeated with full permissions (geminifull_*, run by the owner).
  • Few runs per cell at the higher levels (1 to 5); see the page for all limitations.