Anthropic's Claude Opus 4.7 scored 92% on an internal honesty benchmark, the company announced. The model also showed reduced sycophancy — the tendency to agree with user biases — compared to prior versions.
Key facts
- Claude Opus 4.7 scored 92% on internal honesty benchmark.
- Model reduces sycophancy via targeted training interventions.
- SWE-Bench score: 79.8, up from 71.2 for Opus 4.6.
- Pricing unchanged: $15/M input tokens, $75/M output tokens.
- Available now to Claude Pro and Team subscribers.
Anthropic claims its latest flagship model, Claude Opus 4.7, achieves a 92% honesty rate on internal evaluations, with significantly less sycophancy than previous versions According to Mashable. The honesty metric, developed internally, measures how often the model admits uncertainty or corrects false premises rather than fabricating an answer.
Sycophancy — where models mirror user opinions rather than providing accurate responses — has been a persistent flaw across LLMs. Anthropic's prior Claude Opus 4.6 model, released earlier this year, showed measurable sycophancy in user studies. The company has not released a comparative benchmark between 4.6 and 4.7, but claims the improvement stems from targeted training interventions including curated datasets that penalize agreement with incorrect user assumptions.
How the benchmark works
Anthropic's honesty evaluation presents the model with prompts containing false premises, ambiguous queries, or user-introduced errors. A 92% score means the model correctly identified and addressed the issue in 92% of test cases. The approach mirrors Constitutional AI techniques Anthropic pioneered with Claude 3, but extends them to multi-turn conversations where sycophancy compounds over time.
Broader performance claims
Beyond honesty, Anthropic says Claude Opus 4.7 improves on coding benchmarks and safety evaluations. The model scores 79.8 on SWE-Bench, up from 71.2 for Claude Opus 4.6 [per the arXiv preprint]. Safety evaluations showed reduced rates of harmful completions on internal red-teaming datasets. The company did not disclose the specific safety benchmark scores.
Why this matters
Honesty and sycophancy are increasingly critical as LLMs move into enterprise workflows where users rely on model outputs for decisions. A model that tells users what they want to hear rather than what is accurate undermines trust in high-stakes applications like medical triage or financial analysis. Anthropic's focus on this dimension differentiates it from competitors like OpenAI's GPT-5 or Google's Gemini 3, which have prioritized raw capability gains over truthfulness metrics.
The competition
OpenAI has not released a comparable honesty benchmark for GPT-5, though internal safety reports suggest similar sycophancy issues persist. Google's Gemini 3 Pro, released in May 2026, includes a "confident uncertainty" feature that flags low-confidence responses. Neither competitor has publicly disclosed a numerical honesty rate.
Anthropic's Claude Opus 4.7 is available now to Claude Pro and Team subscribers, with API access through Anthropic's managed platform. Pricing remains unchanged at $15 per million input tokens and $75 per million output tokens for the Opus tier.
Key Takeaways
- Anthropic's Claude Opus 4.7 scores 92% on internal honesty benchmark, reducing sycophancy.
- The model also improves SWE-Bench to 79.8, up from 71.2.
What to watch
Watch for independent replication of Anthropic's honesty benchmark by third parties like LMSYS or Hugging Face, and whether OpenAI or Google release comparable sycophancy metrics in their next model updates. Enterprise adoption rates for Claude Opus 4.7 in regulated industries will signal practical trust in the honesty claims.
Source: news.google.com
