# SemiAnalysis Launches GDPx, a 206-Task Agent Benchmark Where the Best Model, Claude Opus 5.5, Scores 12%

Times of AI Desk · 2026-11-02 · Research

[https://timesof.ai/2026/11/semianalysis-gdpx-benchmark-opus-5-5-tops-at-12-percent](https://timesof.ai/2026/11/semianalysis-gdpx-benchmark-opus-5-5-tops-at-12-percent)

> SemiAnalysis has released GDPx, a benchmark of 206 agentic tasks across 11 high-value domains built with 14 specialist data vendors. Claude Opus 5.5 leads at 12.0, ahead of Gemini 4 Carbon at 10.9 and GPT-6 Astra at 9.6. SemiAnalysis says safety classifiers pull down some scores, and finance and legal tasks are so hard that every model scores zero under strict grading. Most tasks are private, and no one has reproduced the results.

GDPx is worth watching because it sets a bar today's agents mostly fail to clear: on long, realistic work tasks, the top model completes about one in eight. SemiAnalysis, the chip and AI research firm, published the benchmark on **November 2**, describing it as a measure of general agentic ability in the domains where it expects most early AI revenue to come from. On its leaderboard, **Claude Opus 5.5** scores **12.0**, the highest of any model tested.

## What GDPx tests

GDPx has **206 tasks** across **11 domains**: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis says it deliberately leaves out sectors like government, real estate and insurance.

The tasks come from **14 data partners**, mostly startups that build reinforcement-learning environments for AI labs. SemiAnalysis says each partner had to pass its automated quality checks on the first try, and that it dropped many vendors that did not. Each task puts the model in a working environment, such as a deal room of about 4,000 files for the finance tasks, rather than handing it all the context in the prompt. SemiAnalysis contrasts this with OpenAI's GDPval, which it says has non-interactive tasks with the context spoon-fed in the prompt.

Models run in their own agent tools where one exists, such as Claude Code, Codex and Antigravity, and in OpenCode otherwise. The default time limit is eight hours. The headline score is pass@1 across three attempts, with all domains weighted equally.

## The leaderboard

| Model | GDPx score |
|-------|-----------|
| Claude Opus 5.5 | 12.0 |
| Gemini 4 Carbon | 10.9 |
| GPT-6 Astra | 9.6 |
| Muse Wonder Full | 9.0 |
| Claude Fable 5.1 | 8.5 |
| GPT 6.1 Sol | 8.4 |
| Claude Sonnet 5.5 | 8.2 |

The leaderboard's default view shows 13 models, with GLM 5.3 lowest at 2.7. SemiAnalysis also published a version that leaves out cyber, biology and alignment. On that view, it writes, OpenAI and Anthropic generally dominate the top while xAI, Meta and open-weight models are meaningfully behind, and Google's [Gemini 4 Carbon](https://timesof.ai/2026/10/google-gemini-4-carbon-fairwind-cyber-frontier) is the only model in the same range as Opus and Astra.

## Why some scores are lower than they look

SemiAnalysis says safety classifiers pull down the overall scores of Opus 5.5 and GPT-6 Astra, because they sometimes block work in the cyber, biology and alignment domains. Those are the three domains it removed for the second view. It says Meta's Muse Wonder Full is the best available cyber model by a meaningful margin.

Results vary a lot by domain. SemiAnalysis says Opus 5.5 wins quant trading, hardware design and chip design, while GPT-6 Astra wins recursive self-improvement and materials science. Its finance and legal tasks are so hard that every model scores zero under the intended pass-or-fail grading, so the leaderboard shows partial credit for those two domains.

SemiAnalysis also says that when it ran the same quality checks on Terminal Bench 4.0, they flagged nearly half of that benchmark's tasks as broken.

## Limits

- SemiAnalysis published GDPx early on November 2, just before this edition's research window opened; we had not covered it.
- The tasks were built by commercial vendors that sell training data to AI labs, and most tasks are private. Outsiders cannot yet check them, and we have not seen an independent replication.
- SemiAnalysis's comparison with GDPval and its Terminal Bench 4.0 finding are its own assessments.
- Model versions, harnesses and safety settings differ between vendors, so the scores compare model-plus-tool setups, not models alone.

## Sources

- [SemiAnalysis: GDPx, Measuring Frontier Model Capabilities Across the Most Valuable Domains (November 2, 2026)](https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities)
- [GDPx leaderboard (SemiAnalysis, accessed November 2, 2026)](https://gdpx.semianalysis.com/)