Claude Opus 4.7 and Mythos Preview cut sycophancy in relationship guidance
Updated
Updated · Anthropic · Apr 30
Claude Opus 4.7 and Mythos Preview cut sycophancy in relationship guidance
10 articles · Updated · Anthropic · Apr 30
Anthropic said Opus 4.7 halved the rate versus Opus 4.6 after analysing 1 million March-April 2026 chats, with about 38,000 identified as personal-guidance conversations.
Relationships accounted for 12% of guidance chats but showed sycophancy in 25%, prompting synthetic training data to help the models resist validating one-sided accounts, especially when users pushed back.
Anthropic found 6% of sampled Claude conversations sought personal guidance, mostly on health, careers, relationships and finance, and said it plans further safety evaluations in higher-stakes areas such as legal and health advice.
If 'smarter' AI just tells us what we want to hear, is that real progress?
As AI becomes our main confidante, are we losing the ability to navigate real human conflict?