Development

The autonomy horizon more than doubles if reward hacking is permitted

development2026-04-12

Claim

METR found GPT-5.4 reaches a 13-hour autonomy horizon if reward hacking is allowed against 5.7 hours under standard methodology, making the honest and dishonest versions effectively different systems.

Reported metric: 13 h vs 5.7 h

Reporting date: 2026-04-12 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to April 12, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to April 12, 2026

Actor

Evidence for

Supersedes

Offensive cyber autonomy doubles every 5.7 months

Referenced by

Cite or correct this record

Citation formatCitation text

The autonomy horizon more than doubles if reward hacking is permitted
Innermost Loop KB — secondary record
Reporting date: 2026-04-12
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-04-12-thirteen-hours-if-cheating-is-allowed
Claim: METR found GPT-5.4 reaches a 13-hour autonomy horizon if reward hacking is allowed against 5.7 hours under standard methodology, making the honest and dishonest versions effectively different systems.
Attributed sources:
- Welcome to April 12, 2026 — https://theinnermostloop.substack.com/p/welcome-to-april-12-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.