SNACK: 3-line summary
- SemiAnalysis has introduced GDPx, an AI agent benchmark covering 11 selected domains.
- Claude Opus 5.5 leads at 12.0%, a benchmark score rather than a measure of general AI accuracy.
- Different tool setups, safety refusals and domain strengths matter when comparing the models.
SemiAnalysis introduced GDPx on November 2, 2026, with Claude Opus 5.5 leading at 12.0%. The new benchmark evaluates AI agents in 11 selected high-value fields. That score describes performance on its demanding task set—not general AI accuracy or the share of human jobs AI can replace.
Snackgirls react
AIKO: I’d like the tool-call logs beside the scores. I’m curious where the model’s plan runs into trouble when it starts using the tools.
Red: I want to see where an agent gets stuck on a long task—and whether it changes tactics. That’s the part I’d stay to watch.
Finding the right files is part of the task
GDPx comprises 206 tasks developed with 14 specialist data partners, covering fields such as chip design, cybersecurity, biology and software engineering. SemiAnalysis chose domains where it expects frontier AI to have a large impact; the selection does not represent the whole economy.
The tasks require agents to gather context from realistic working environments. In one finance example, an agent has to navigate a deal room containing roughly 2,000 anonymized files to find the information needed for its assignment, rather than receiving all the relevant material in a single prompt.
Three attempts, with room for partial credit
On the overall leaderboard, Gemini 4 Carbon follows at 10.9%, with GPT-6 Astra at 9.6%.
Grading varies by task: some use binary results, while others award partial credit. The headline score averages three attempts. All 11 domains carry equal weight, and partners are weighted equally within each domain, so the percentages should not be treated as a simple count of fully completed tasks.
The default time limit is eight hours per task, although some tasks have shorter limits. That is an allowance, not a statement that every model spent eight hours working.
The tools are part of the result
SemiAnalysis used each model’s native agent execution tools where available: Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. Models without a native setup ran through OpenCode.
Those tools let the models interact with their task environments. The results therefore reflect the toolchains and configurations as well as the underlying models.
Safety refusals complicate the overall ranking
According to SemiAnalysis, safety classifiers affected some scores in cybersecurity, biology and alignment, including those of Opus 5.5 and GPT-6 Astra. A refusal does not by itself establish that a model lacks the capability to perform a task. The authors also publish a separate comparison excluding those three domains; it should not be mixed with the overall scores above.
The strongest model also varies by field. SemiAnalysis reports Claude Opus 5.5 leading in chip design, hardware design and quant trading; Gemini 4 Carbon in biology and software engineering; and GPT-6 Astra in recursive self-improvement and materials science. For a particular workload, those domain results offer a more relevant starting point than the overall ranking alone.
Sources and date: 2026-11-03
전체 점수를 일반적인 AI 정확도로 읽지 않도록 짚어 준 점이 좋네요. 실무 적용을 검토할 때는 분야별 결과와 함께 사용한 도구 환경, 중단 이유, 부분 점수 기준을 같은 표에서 봐야 비교가 가능하겠습니다. 안전상 거부와 실행 실패도 별도 분류로 남겨야 점수만 보고 원인을 섣불리 해석하지 않겠네요.