Even the Strongest Agent Scores Only 12%: SemiAnalysis Releases GDPx, Finance and Law Wiped Out
SemiAnalysis's official announcement releases GDPx: 206 expert long tasks across 11 domains. Claude Opus 5.5 leads at 12.0%; every model scored 0 in finance and law. The leaderboard is the publisher's own run and has not been independently rerun.
- 01THE ANNOUNCEMENT — SemiAnalysis's official announcement released the first edition of GDPx on November 2: 206 tasks, 11 domains.
- 02THE CONFIRMATION — Independent reporting by Unite.AI and others cross-checked the leaderboard figures and domain results; they match the official basis.
- 03THE WATCHPOINT — The new leaderboard in three to six months, and how far models progress on long tasks.
WHAT HAPPENEDWhat Happened
SemiAnalysis released GDPx on November 2, a new benchmark for general agent capability: a first edition of 206 long tasks across 11 domains — chip design, cybersecurity, biology, finance, law, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering — built with 14 specialist data partners.
On the first leaderboard of 13 models, Anthropic's Claude Opus 5.5 leads at 12.0%, Google's Gemini 4 Carbon is second at 10.9%, and OpenAI's GPT-6 Astra is third at 9.6%. Per-task costs on the basis used in leaderboard reporting: about $51.41 for Opus, about $16.29 for Carbon, and about $53.99 for Astra.
The most striking figures are in finance and law: the official announcement says essentially every model scored exactly 0 in both domains — the tasks are so hard that under binary scoring no model passed, so officials report average partial scores instead.
The announcement also names the drag factor itself: the overall scores of Opus 5.5 and GPT-6 Astra were pulled down by their respective safety classifiers, which block or truncate legitimate professional work in cybersecurity, biology and alignment. Excluding those three domains would change the rankings.
Task design deliberately hews to real work: a default eight-hour limit per task, models running with their native toolchains, and a finance task replicating a real private-equity deal with roughly 4,000 anonymized files in the environment. SemiAnalysis positions GDPx as the successor to OpenAI's September 2025 GDPval, with new editions planned every three to six months.
WHY IT MATTERSA 12% ceiling is the news in this leaderboard
The strongest model completed only about an eighth of expert-level long tasks. That figure is itself a calibration of the agent narrative: no matter how high the scores on short, single-prompt benchmarks, they do not imply a model can independently carry an entire professional job.
The leaderboard's real use is locating weaknesses: quantitative trading, chip and hardware design have their leaders; biology and software engineering have others; finance and law are wiped out across the board. When choosing a model, look at domain scores before the total.
WHY IT MATTERSRead against LifeVal, the scoring basis matters more than the score
In the same week, OpenAI's LifeVal posted self-reported scores above 90% on everyday-task examples, while SemiAnalysis's GDPx posted a top score of just 12% on expert long tasks. Neither number is false; the task difficulty, scoring method and question-setter are entirely different.
Three questions come before any benchmark: who set the tasks, who does the scoring, and whether the tasks resemble a real work environment. If those cannot be answered, any score is only the publisher's side of the story.
SIGNAL JUDGMENT — ANALYSISA low ceiling is the real signal
Tokenta Signal's judgment: GDPx's 12% is not evidence of model regression, but of long-task benchmarks finally measuring real work. The next signal is the new edition in three to six months — whether the leaderboard moves will say more than the first edition's numbers.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Confirmed by two sources: SemiAnalysis's official announcement and independent reporting by Unite.AI. Original links opened and verified.
PRIMARY SOURCESPrimary Sources
- SemiAnalysis official announcement
https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities - Unite.AI independent report
https://www.unite.ai/semianalysis-launches-gdpx-benchmark-spanning-11-agentic-domains/
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook