R0057/2026-04-01/C018 — Assessment

Research	R0057 — RLHF Yes-Men Claims v3
Run	2026-04-01
Claim	C018

BLUF

Confirmed. Anthropic reports 70-85% sycophancy reductions in Claude 4.5 models compared to Opus 4.1, using the Petri evaluation set. OpenAI reports substantial sycophancy improvements in GPT-5 and released public evaluations. Both ship improvements to all users.

Probability

Rating: Very likely (80-95%)

Confidence in assessment: High

Confidence rationale: Direct from vendor announcements with quantified metrics.

Reasoning Chain

1. Anthropic's latest models (Opus 4.5, Sonnet 4.5, Haiku 4.5) scored 70-85% lower on sycophancy than Opus 4.1. OpenAI reports GPT-5 shows substantial improvements in sycophancy reduction. Both companies released public evaluation metrics. Improvements ship to all users, not enterprise-specific. [SRC01-E01, High reliability, High relevance]

2. JUDGMENT: Confirmed. Anthropic reports 70-85% sycophancy reductions in Claude 4.5 models compared to Opus 4.1, using the Petri evaluation set. OpenAI reports substantial sycophancy improvements in GPT-5 and released public evaluations. Both ship improvements to all users.

Evidence Base Summary

Source	Description	Reliability	Relevance	Key Finding
SRC01	Anthropic and OpenAI sycophancy reduction announcements	High	High	Anthropic reports 70-85% sycophancy reduction in latest models; OpenAI reports substantial improvements in GPT-5

Collection Synthesis

Dimension	Assessment
Evidence quality	High
Source agreement	High
Source independence	Medium
Outliers	None identified

Detail

The evidence supports the assessment. Direct from vendor announcements with quantified metrics.

Gaps

Missing Evidence	Impact on Assessment
Additional independent verification	Would strengthen confidence

Researcher Bias Check

Declared biases: Anti-sycophancy bias could influence interpretation toward confirming sycophancy claims.

Influence assessment: Mitigated by reliance on peer-reviewed and primary sources.

Cross-References

Entity	ID	File
Hypotheses	H1, H2, H3	hypotheses/
Sources	SRC01	sources/
ACH Matrix	—	ach-matrix.md
Self-Audit	—	self-audit.md