SemiAnalysis GDPx benchmark shows AI agents plateau below 12 percent on knowledge work
Source published ·Modelwire updated
Original coverage: The Decoder ↗·How Modelwire adds context
The development
SemiAnalysis has released GDPx, a benchmark designed to measure AI agent performance on realistic knowledge work tasks spanning chip design, quantitative trading, and related domains. The results reveal a significant capability gap: the highest-performing models, Claude Opus 5.5 and Gemini 4 Carbon, achieved only 12 percent accuracy. This finding underscores that despite recent advances in reasoning and planning, current AI agents remain far from autonomous execution of complex professional workflows. The benchmark's methodology has drawn scrutiny for limited task diversity per domain and inconsistent testing protocols, raising questions about whether the gap reflects genuine limitations or measurement artifacts. For enterprises evaluating agentic AI for high-stakes work, the results suggest deployment timelines remain longer than hype cycles suggest.
Modelwire’s AI-generated summary of coverage from The Decoder.
Modelwire analysis
Skeptical readOur AI-generated reading of the wider context and the next developments to watch.
SemiAnalysis hasn't disclosed whether the 12% floor holds across repeated runs or task variants, leaving open the possibility that results reflect benchmark brittleness rather than agent incapacity. The summary notes methodological scrutiny but doesn't specify what tasks were excluded or why testing protocols varied.
This echoes a pattern from the past week. CloudOpsBench (Nov 4) identified reasoning as the bottleneck in DevOps workflows, but the 'When Agents Fail' paper (Oct 23) showed that benchmarks routinely conflate model weakness with harness design failure. SlackBench (Nov 3) went further, arguing that isolated synthetic tasks miss the coordination and context-synthesis demands of actual work. GDPx follows the same isolated-task template. The real question isn't whether 12% is low, but whether a benchmark measuring chip design and trading in lab conditions tells us anything about agent performance in messy, collaborative environments where information scatters across chat, code, and evolving decisions.
If SemiAnalysis releases task-level breakdowns showing Claude and Gemini perform similarly across chip design, trading, and other domains, that suggests a general reasoning floor. If performance diverges sharply by domain, the 12% aggregate masks that some agents are already viable for specific workflows, and the benchmark is measuring breadth, not depth. Watch whether the same models score higher on SlackBench's collaborative tasks than on GDPx's isolated ones within 60 days; if so, the gap is partly about benchmark design, not capability.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Framework separates model failures from harness design flaws in agents
A new evaluation framework addresses a critical blind spot in agentic AI benchmarking: the joint performance of language models and their execution harnesses obscures which component actually fails. The paper introduces five diagnostic measures spanning completion reliability, model-harness compatibility, cross-model generalization, interaction-specific residuals, and token efficiency. This matters because current benchmarks conflate model capability with…
MentionsSemiAnalysis · GDPx · Claude Opus 5.5 · Gemini 4 Carbon
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. The Decoder originally reported this story as “No AI agent scores above 12 percent on SemiAnalysis' new real-world work benchmark”. The full content lives on the-decoder.com. If you’re a publisher and want a different summarization policy for your work, see our takedown page.