# GDPx benchmark: best model, Claude Opus 5.5, scores only 12%

> SemiAnalysis tests frontier models on 206 expert tasks in 11 fields, and finance and legal work mostly goes unsolved.

URL: https://claudegpt.si/news/gdpx-benchmark-opus-5-5-12-percent  
Published: 2026-11-02T06:53:07.560Z  
Author: ClaudeGPT Si Newsroom  
Section: Research  
Topics: SemiAnalysis, Benchmarks, Anthropic, Claude, AI agents  
Publisher: ClaudeGPT Si

## Key takeaways

- SemiAnalysis launched GDPx on November 2 with 206 tasks from 14 data partners across 11 domains.
- Claude Opus 5.5 leads with a 12% score, and nearly every model scores 0 on binary-graded finance and legal tasks.
- SemiAnalysis says safety classifiers lower some models' scores in cyber, biology and alignment, so rankings change without them.

SemiAnalysis released GDPx on November 2, a benchmark of how well frontier AI models do expert work in 11 fields, and the best model, Claude Opus 5.5, scores 12%. In the [announcement](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities), the authors call GDPx "a true frontier-level benchmark", meaning the current models are nowhere near done with it.

It matters because most public benchmarks are now close to full marks. A test where the leader fails 88% of the time shows where the gaps are, and the gaps differ a lot by field.

## What GDPx tests

GDPx has 206 tasks written with 14 specialized data partners. The 11 domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (an AI improving AI systems), hardware design, materials science, alignment and software engineering. SemiAnalysis left out government, real estate and insurance.

The headline score is pass@1 averaged over three attempts, which means the share of tasks solved on the first try. Each domain counts equally. Grading is binary or partial depending on the domain. Tasks run in realistic environments, with an 8-hour limit by default. Each model uses its maker's own agent harness where one exists, such as Claude Code for Opus 5.5 and Codex for GPT-6 Astra, and OpenCode otherwise.

## Who scores where

SemiAnalysis says the picture changes once you drop three domains, cybersecurity, biology and alignment, because safety classifiers lower some models' scores there. Without them, OpenAI and Anthropic lead, and Gemini 4 Carbon is the only other model in the same range. SpaceXAI, Meta and open-weight models trail.

According to SemiAnalysis:

- Opus 5.5 is the best all-around model and wins quant trading, hardware design and chip design.
- GPT-6 Astra wins recursive self-improvement clearly and leads materials science.
- Gemini 4 Carbon is the best biology model and edges out Opus in software engineering.
- Muse Wonder Full is the best cyber model by a meaningful margin.
- On cost against performance, GPT-6 Luna and DeepSeek lead at the low end, Opus at the high end and OpenAI models in the middle.

## Where models fail

Finance and legal show the problem most clearly. In finance, nearly every model scores 0 under binary grading. On one private-equity task built from about 4,000 files, no model solved it in 25 attempts. The best single run passed 8 of 11 checks.

In legal, no model produced a usable contract redline across 20 rollouts. Partial scores averaged 44% for Opus 5.5, 25% for GPT-6 Astra, 2% for Gemini 4 Argon and 0.5% for GLM 5.3. The best single run reached 57%.

In hardware design, none of 12 rollouts on a compressor-housing task produced a verified full pass.

## Why SemiAnalysis built it

The authors argue that GDPval, OpenAI's earlier work-task benchmark released September 25, 2025, has weak spots. They point to single-shot tasks with no environment, and rubric grading where human raters agreed only 71% of the time. GDPx instead uses deterministic or verifiable graders where possible and domain experts as task authors.

## What to keep in mind

This is a new benchmark from one firm, with partner-written tasks. The authors say partners could game it with enough iteration, so they required each task to pass automated quality checks on the first try. They also say performance is "spiky": every leading model is strong in some domains and weak in others. Gemini 4 Carbon was not yet generally available when the post went up.

Several domains have public tasks, and the full comparisons are on the [GDPx site](https://gdpx.semianalysis.com). For another look at how agent benchmarks can mislead, see our report on [MCP versus CLI scores](https://claudegpt.si/news/mcp-vs-cli-agents-study).

## Sources

- [GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities), SemiAnalysis
- [GDPx](https://gdpx.semianalysis.com), SemiAnalysis