Unverified50% confidenceBenchmarkExact time
在GPT-4发布初期，主流模型在GPQA上的准确率约为35-40%，人类领域专家基准线约为65%，随机猜测为25%
1
Sources
50%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
OpenAI's Head of Evals: We've Been Consistently Underestimating What AI Models Can Really Do
bilibiliGelai_AI7/8/2026
Related Claims
UnverifiedClaude Opus 5在智能体搜索BrowseComp上得分30.8%，与GPT系列基本持平；工具增强多学科推理64.7%；生物学人类可解问题上高达90.1%72% similarUnverifiedOn the coding index (reasoning mode) in third-party evaluations by Artificial Analysis, GPT-5 Mini High scored 9 points higher than Haiku 4.571% similarUnverifiedGPQA is a graduate-level question-answering benchmark introduced in 2023; even with the aid of a search engine, non-experts achieve a correct answer rate of only about 30%.71% similarUnverifiedGPQA由纽约大学2023年发布，题目设计为即便借助搜索引擎普通人正确率也仅约34%71% similarUnverifiedEarly GPT-4 had a pass rate of only 1.7% on SWE-bench, Claude 3 Opus approximately 4.8%, and it was not until mid-2024 that the 20% mark was surpassed.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/362872API
curl https://kongchang.com/api/v1/knowledge/claims/362872MCP
get_claim(id=362872)