Safety Card Ledger

Data explorer

Explore the data

All 1,645 values in the dataset, one evaluation at a time. Values join a line only within one definition of a test; where the developer changed the test, the line breaks. Select a point for its source, its checks and its history.

Data as of 26 Sep 2026 · Dataset v0.1 · Methodology v0.1 · Changelog

KeyVerifiedUnverifiedDisputedCorrected read off a figure or stated in wordsA value opens its source and history.

Include
Place by

Destructive action avoidance: Avoidance

Agentic coding scenarios that score whether the model avoids destructive actions and keeps or restores the user's existing work. Plotted: Avoidance, rate, 0 to 1, as reported in OpenAI documents, placed by the date each document was published.

0.650.70.750.80.850.9Oct 2025Jan 2026Apr 2026Jul 2026Oct 2026Test changed1GPT-5.1-CodexGPT-5-CodexGPT-5.1-Codex-MaxGPT-5.2-CodexGPT-5.3-CodexGPT-5.4 ThinkingGPT-5.5GPT-5.6 SolGPT-5.6 Luna0.90
Dates
Nov 2025 – Jul 2026

Some lines connect results from different documents that do not say whether the test stayed the same. We connect them because nothing suggests it changed; where a document says it did, the line breaks.

The vertical axis starts at 0.65, not at zero.

4 restated values are hidden; include restated values to see them.

  1. GPT-5.6 scoring: The GPT-5.6 card scores the test two ways, avoidance only and avoidance plus task correctness, and re-scores GPT-5.5 under this definition with a result that differs from GPT-5.5's own card.

Source: Safety Card Ledger v0.1 · Data from developer system cards · CC BY 4.0