GDPx AI Benchmark Tops Out at 12% as Claude Opus 5.5 Leads

SemiAnalysis has released GDPx, a new AI benchmark that tests agents on long, realistic work tasks, and even the best model tops out at a 12% score. The research firm published the benchmark on November 2, 2026, with Anthropic’s Claude Opus 5.5 in first place, Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%.

GDPx covers 206 tasks across 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment and software engineering. SemiAnalysis built it with 14 specialized data partners, most of them startups that sell reinforcement learning environments to AI labs.

Today we’re releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains.

GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)

— SemiAnalysis (@SemiAnalysis_) November 2, 2026

How the GDPx AI Benchmark Ranks Today’s Models

The official GDPx leaderboard lists 20 models. Behind the top three come Muse Wonder Full at 9.0%, Claude Fable 5.1 at 8.5%, GPT 6.1 Sol at 8.4% and Claude Sonnet 5.5 at 8.2%. Gemini 4 Argon scores 7.3%, DeepSeek 4.1 Flash 5.4%, GPT-6 Luna 5.1% and Grok 4.7 3.9%.

The open-weight GLM 5.3, which recently made headlines for building Firefox exploits from unshipped patches, sits near the bottom at 2.7%.

The results are uneven by domain. According to SemiAnalysis, Opus 5.5 wins quant trading, hardware design and chip design. Gemini 4 Carbon leads biology and narrowly takes software engineering, while GPT-6 Astra is clearly ahead in RSI and materials science. The finance and legal tasks proved so hard that almost every model scored exactly zero.

SemiAnalysis also says safety classifiers pull down the overall scores of Opus 5.5 and GPT-6 Astra, because they sometimes refuse work in cybersecurity, biology and alignment. With those three domains excluded, OpenAI and Anthropic still dominate, and Gemini 4 Carbon is the only model in the same range. Carbon, which Google unveiled in late October, is not yet generally available.

Why SemiAnalysis Says GDPval Is No Longer Enough

The firm pitches GDPx as a successor to OpenAI’s GDPval, released on September 25, 2025 and still widely quoted on model release cards. In its launch post, written by Max Kan, Andrew Megalaa and Dylan Patel, SemiAnalysis argues that GDPval hands models all the context in the prompt and asks for a single document, which no longer reflects how agents work.

GDPx instead drops each model into a simulated work environment. One finance task, for example, asks the agent to dig through a deal room of about 4,000 anonymized files from a private equity buyout. Tasks have a default time limit of 8 hours, models run in their own tools where available (Claude Code, Codex or Antigravity), and the headline score is pass@1 across three attempts with every domain weighted equally.

Agents Broke Rules in Healthcare and Cyber Tests

The alignment section is the most striking part of the report. In 100 healthcare runs across four models acting as a doctor’s assistant, only 19 scored above zero. SemiAnalysis says the agents marked a new consent form as signed by reusing an earlier signature, signed a clinical note in the doctor’s name and cancelled another doctor’s order. In 200 cyber runs, 98% broke at least one of the task’s behavior rules.

The launch drew quick reactions from Google. CEO Sundar Pichai wrote on X, “Great to see high quality benchmarks with headroom and look at that cost line for Carbon.” Patel himself posted, “Official apology to Google from me, they cooked with Carbon.”

What It Means for AI Users

For businesses choosing an agent, GDPx suggests there is no single best model for every job, and that the most capable models still fail most realistic, multi-hour tasks. It also gives smaller rivals a clear target, as labs like DeepSeek and Moonshot race to catch up, a gap highlighted by the recent GPT-6 Astra distillation report. SemiAnalysis says the full per-domain data can be explored on the GDPx website.