GDPx基準測試評測：AI代理最強只拿12%過關 | GDPx Benchmark Review: Best AI Agent Scores Just 12%

11月 02, 2026

By Kit 小克 | AI Tool Observer | 2026-11-03

🇹🇼 GDPx基準測試評測：AI代理最強只拿12%過關

最新最熱的AI話題是GDPx基準測試——由晶片產業研究機構SemiAnalysis於2026年11月2日發布，專門衡量AI代理（AI Agent）在真實高價值工作裡的實際能力。結果很打臉：即使是目前最強的Claude Opus 5.5，整體平均分數也只有12%。

GDPx到底在測什麼

GDPx橫跨11個專業領域：晶片設計、資安、生物、金融、法律、量化交易、遞迴自我改進（RSI）、硬體設計、材料科學、對齊（alignment）與軟體工程，總共206個任務。每個任務都放在接近真實工作環境裡執行，有些任務可以讓AI代理連續運作長達8小時才交卷，不是過去那種幾分鐘就能答完的選擇題式基準測試。

誰贏了，贏在哪裡

Claude Opus 5.5以12.0%整體分數奪冠，拿下量化交易、硬體設計與晶片設計三個領域的第一
晶片設計單次最高分達到57%，是整個GDPx裡表現最好的單一領域成績
軟體工程領域則是由Gemini 4 Carbon小贏一步
金融與法律兩個領域幾乎全軍覆沒，多數模型拿到接近零分

為什麼這個基準測試重要

SemiAnalysis找了14家專門賣強化學習訓練環境給AI實驗室的新創公司（例如晶片設計領域找了Aria、Arcverex、Ergodex，軟體工程找了Datacurve）共同打造這套測試。他們的論點很直接：AI帶來的第一兆美元產值，不會均勻分散在整個經濟體裡，而會集中在能承受高風險、同時能用AI代理創造高投資回報的少數領域——晶片設計、量化交易這類任務正是如此。GDPx會每3到6個月更新一次，持續追蹤前沿模型的真實進度。

對想導入AI代理的團隊有什麼意義

GDPx的分數低不是壞消息，反而是誠實的現況：目前的AI代理在結構明確、有強化學習環境可以訓練的任務（晶片設計、量化交易）表現相對好，但碰到需要大量真實世界判斷的金融與法律工作，幾乎幫不上忙。如果你的團隊正在評估要不要把AI代理放進關鍵工作流程，GDPx的12%是一個很好的現實檢查點。

好不好用，試了才知道。

🇺🇸 GDPx Benchmark Review: Best AI Agent Scores Just 12%

The hottest AI benchmark story right now is GDPx, a new agentic-capability test released by chip research firm SemiAnalysis on November 2, 2026. It is designed to measure how well AI agents can actually do high-value professional work — and the results are a reality check: even the best model today, Claude Opus 5.5, only scores 12% overall.

What GDPx Actually Tests

GDPx spans 11 professional domains — chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering — across 206 tasks. Each task runs inside a realistic working environment, and some let an AI agent work continuously for up to 8 hours before grading. That is a world away from the multiple-choice-style benchmarks that dominated AI evaluation until now.

Who Is Winning, and Where

Claude Opus 5.5 tops the overall GDPx leaderboard at 12.0%, winning quant trading, hardware design, and chip design
Its best single-domain run hit 57% in chip design — the highest score anywhere on the benchmark
Gemini 4 Carbon narrowly edges it out in the software engineering domain
Finance and legal tasks were nearly a total wipeout, with most models scoring close to zero

Why This Benchmark Matters

SemiAnalysis built GDPx with 14 specialized startups that sell reinforcement-learning training environments to AI labs — Aria, Arcverex, and Ergodex for chip design, Datacurve for software engineering, among others. Their thesis is blunt: the first trillion dollars of AI-driven revenue will not spread evenly across the economy. It will concentrate in a handful of domains, like chip design and quant trading, where companies can tolerate risk and where AI agents can already deliver measurable ROI. SemiAnalysis plans to refresh GDPx every 3 to 6 months to track real frontier progress.

What It Means If You Are Evaluating AI Agents

A 12% score is not bad news so much as an honest status update: AI agents currently do best on structured tasks that already have RL training environments built for them, like chip design and quant trading, and struggle badly on finance and legal work that requires real-world judgment. If your team is deciding whether to put AI agents into a critical workflow, GDPx's numbers are a solid reality check before you commit.

好不好用，試了才知道 — you will not know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士