SemiAnalysis 发布 GDPx 基准测试，评估 AI 智能体处理专家工作的能力Chinese only
GDPx 覆盖 11 个专业领域，要求智能体自行检索工作区文件来完成任务。测试中排名第一的 Claude Opus 5.5 平均得分仅为 12%；在改进交易策略的 111 次尝试中，模型有 70 次表现不如原策略，显示这些任务仍有明显难度。
1 developments2 reports1 sources1 verified facts
Verified facts 1
SemiAnalysis has released GDPx, a benchmark testing AI agents on expert work across 11 domains.
SemiAnalysis发布的GDPx基准测试显示，Claude Opus 5.5平均得分为12%，Gemini 4 Carbon为10.9%；OpenAI的GPT-6 Astra在AI研究和材料科学任务中得分最高。
Event developments 1
SemiAnalysis releases GDPx benchmark for AI agents doing expert work across 11 domains
GDPx requires agents to search workspaces rather than receive context in the prompt. Anthropic’s Claude Opus 5.5 averages 12%, followed by Google’s Gemini 4 Carbon at 10.9%; OpenAI’s GPT-6 Astra leads on AI research and materials science tasks. Every model scores zero on finance and law tasks under pass-fail grading.
View source reports · 1 sources · 2 reports
- Rohan Paul (@rohanpaul_ai) · SemiAnalysis releases GDPx benchmark for AI agents doing expert work across 11 domains
- Rohan Paul (@rohanpaul_ai) · SemiAnalysis releases GDPx benchmark for AI agents doing expert work
All reports 2
SemiAnalysis releases GDPx benchmark for AI agents doing expert work
GDPx covers 11 expert domains and requires agents to search workspace files to complete tasks. The top-ranked Claude Opus 5.5 averaged just 12%; in 70 of 111 attempts to improve a trading strategy, models performed worse than the original.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)
SemiAnalysis releases GDPx benchmark for AI agents doing expert work across 11 domains
GDPx requires agents to search workspaces rather than receive context in the prompt. Anthropic’s Claude Opus 5.5 averages 12%, followed by Google’s Gemini 4 Carbon at 10.9%; OpenAI’s GPT-6 Astra leads on AI research and materials science tasks. Every model scores zero on finance and law tasks under pass-fail grading.
1 more source reported this
- Rohan Paul (@rohanpaul_ai)