LessWrong posts by zvi
LessWrong posts by zvi
zvi
“Claude Opus 4.6: System Card Part 1: Mundane Alignment and Model Welfare” by Zvi
55 minutes Posted Feb 9, 2026 at 9:36 pm.
Thinking mode gives better [...] ---Outline:
A Three Act Play
Safety Not Guaranteed
Pliny Can Still Jailbreak Everything
Transparency Is Good: The 212-Page System Card
Mostly Harmless
Mostly Honest
Agentic Safety
Prompt Injection
Key Alignment Findings
Behavioral Evidence (6.2)
Reward Hacking and 'Overly Agentic Actions'
Metrics (6.2.5.2)
All I Did It All For The GUI
Case Studies and Targeted Evaluations Of Behaviors (6.3)
Misrepresenting Tool Results
Unexpected Language Switching
The Ghost of Jones Foods
Loss of Style Points
White Box Model Diffing
0:00
55:33
Download MP3
Show notes
Claude Opus 4.6 is here. It was built with and mostly evaluated by Claude.
Their headline pitch includes:
1M token context window (in beta) with State of the art retrieval performance.
Improved abilities on a range of everyday work tasks. Model is improved.
State of the art on some evaluations, including Terminal-Bench 2.0, HLE and a very strong lead in GDPval-AA.
Claude Code now has an experimental feature called Agent Teams.
Claude Code with Opus 4.6 has a new fast (but actually expensive) mode.
Upgrades to Claude in Excel and the release of Claude in PowerPoint.
Other notes:
Price remains $5/$25, the same as Opus 4.5, unless you go ultra fast.
There is now a configurable ‘effort’ parameter with four settings.
Refusals for harmless requests with rich context are down to 0.04%.
Data sources are ‘all of the above,’ including the web crawler (that they insist won’t cross CAPTCHAs or password protected pages) and other public data, various non-public data sources, data from customers who opt-in to that and internally generated data. They use ‘several’ data filtering methods.
Thinking mode gives better [...] ---Outline:(
03:45) A Three Act Play(04:57) Safety Not Guaranteed(10:53) Pliny Can Still Jailbreak Everything(12:48) Transparency Is Good: The 212-Page System Card(13:53) Mostly Harmless(17:45) Mostly Honest(19:01) Agentic Safety(20:27) Prompt Injection(23:07) Key Alignment Findings(33:48) Behavioral Evidence (6.2)(38:40) Reward Hacking and 'Overly Agentic Actions'(40:37) Metrics (6.2.5.2)(42:40) All I Did It All For The GUI(43:58) Case Studies and Targeted Evaluations Of Behaviors (6.3)(44:19) Misrepresenting Tool Results(45:09) Unexpected Language Switching(46:12) The Ghost of Jones Foods(47:54) Loss of Style Points(48:54) White Box Model Diffing(49:13) Model Welfare ---
First published:
February 9th, 2026
Source:
https://www.lesswrong.com/posts/sWsSncqMLKyGZA9Ar/claude-opus-4-6-system-card-part-1-mundane-alignment-and
---
Narrated by TYPE III AUDIO.
---Images from the article:Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.