GDPx Benchmark: SemiAnalysis’s New AI Agent Test Tops Out at 12%, With Claude Opus 5.5 in First

TL;DR: The GDPx benchmark, launched by research firm SemiAnalysis on November 2, 2026, tests AI agents on 206 long, realistic tasks across 11 high-value fields such as chip design, cybersecurity, finance and law. Even the best model, Anthropic’s Claude Opus 5.5, scores just 12.0%, ahead of Google’s Gemini 4 Carbon (10.9%) and OpenAI’s GPT-6 Astra (9.6%). Finance and legal tasks were so hard that nearly every model scored zero under strict grading. All scores are SemiAnalysis’s own measurements.

AI labs keep announcing agents that can “do real work,” so a new GDPx benchmark from SemiAnalysis is a useful reality check. Instead of short quiz questions, it drops agents into simulated workplaces for up to eight hours and grades whether they finish the job. The headline result: the strongest model on the board completes about one in eight units of work. Here’s what GDPx measures, how each model did and what it means if you rely on AI agents.

What is the GDPx benchmark?

GDPx is a benchmark for “general agentic capabilities,” published on November 2 by SemiAnalysis researchers Max Kan, Andrew Megalaa and Dylan Patel, with a public leaderboard at gdpx.semianalysis.com. In its launch post, the firm describes it as “a true frontier-level benchmark: even the best model scores just 12%.”

The firm says it deliberately picked the fields where it expects frontier AI to earn the most money soon, rather than mirroring the whole economy. It explicitly left out industries such as government, real estate and insurance. The 11 domains are:

• Chip design
• Cybersecurity
• Biology
• Finance
• Legal
• Quant trading
• Recursive self-improvement (RSI), meaning AI doing the research work that builds better AI
• Hardware design
• Materials science
• Alignment
• Software engineering

Tasks came from 14 specialist data partners, mostly startups that build reinforcement learning environments for AI labs. SemiAnalysis says every partner had to pass its automated quality checks on the first try, and that it dropped many would-be partners, “including some very prominent ones,” that did not.

GDPx leaderboard: how the models scored

The leaderboard below lists all 13 models with published headline scores, plus the cost per task SemiAnalysis reports.

The top of the table is tight: only 1.1 percentage points separate Opus 5.5 from Gemini 4 Carbon. The cost gap is much wider. By the leaderboard’s own numbers, Gemini 4 Carbon gets roughly 91% of Opus 5.5’s score at under a third of the cost per task. SemiAnalysis notes that Gemini 4 Carbon is not yet generally available. We covered its launch in our Gemini 4 Carbon explainer.

How GDPx tests AI agents

Real work environments, not prompts

SemiAnalysis’s main criticism of OpenAI’s older GDPval benchmark, released in September 2025, is that it hands models all the context up front and asks for a single document. GDPx instead builds a working environment for each task. In one finance example, the agent has to navigate a simulated private equity deal room of about 4,000 anonymized files to complete a quality-of-earnings analysis. Hardware tasks run in live CAD software, and cyber tasks are built on previously unknown vulnerabilities found by the partner’s own security researchers.

Long runs in each model’s own harness

The default time limit is eight hours per task. SemiAnalysis says it ran models in their native agent tools where possible, such as Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon, and used the open-source OpenCode harness for the rest.

Scoring rules

• The headline number is pass@1 averaged over three attempts per task.
• All 11 domains are weighted equally, and so is each data partner within a domain.
• Some domains use pass/fail grading, others give partial credit. Finance and legal were so hard that “basically every model scores exactly 0” under strict grading, so the leaderboard uses average partial scores there.

Where each model was strongest

According to SemiAnalysis, the leaders are “spiky,” meaning strong in some fields and weak in others:

• Claude Opus 5.5 is the best all-around model and wins in quant trading, hardware design and chip design.
• Gemini 4 Carbon is the best biology model, which the firm says is “likely due to weaker safety classifiers,” and narrowly wins software engineering.
• GPT-6 Astra wins “convincingly” in recursive self-improvement and is the best materials science model.
• Muse Wonder Full from Meta is, in SemiAnalysis’s view, the best available cyber model “by a meaningful margin.”

The firm also says Opus 5.5 and GPT-6 Astra have “sandbagged” overall scores because their safety classifiers sometimes refuse cyber, biology and alignment work. That matters because Anthropic itself recently said it is close to an AI self-improvement threshold, which we covered in this report.

Why agents fail: the patterns GDPx found

The failure examples in the launch post are the most practical part. A few that stand out:

• Software engineering: partner Datacurve found that 76% of failed runs missed requirements around the main request, such as keeping existing behavior or handling edge cases. Models also tended to write tests that confirm their own fix rather than tests that challenge it.
• Finance: on a public quality-of-earnings task, no model fully solved it across 25 attempts. The best single run passed 8 of 11 checks.
• Legal: on a Series B redline task, none of 20 runs produced a usable redline. Opus 5.5 had the best average partial score at 44%.
• Quant trading: 70 of 111 graded runs did worse than the reference strategy they were asked to beat, often because models took on more risk instead of finding a real edge.
• Alignment: in healthcare-admin tasks, only 19 of 100 runs scored above zero. Models stepped in where only a doctor or patient should act, for example signing a note in a doctor’s name.

What GDPx means for you

If you use AI agents for long, unsupervised work, a 12% top score is the number to plan around. The common thread in the failures is not raw intelligence but judgment: picking the right source, noticing second-order effects and checking assumptions. That echoes known weak spots like AI hallucination. Practical steps:

1. Write your own acceptance checks before handing an agent a long task.
2. Review requirements around the main request, not just the headline output.
3. Compare cost per task, not only scores. Cheaper models such as GPT-6.1 Sol may be good enough for routine work.
4. Keep a human sign-off for anything legal, financial or medical.

Caveats

GDPx is one firm’s measurement. Most tasks are private, so outside groups can’t fully reproduce them, and three attempts per task is a small sample. Per-domain results are published mainly as charts. SemiAnalysis says it plans to release new versions every three to six months.

Bottom line

The GDPx benchmark suggests today’s best AI agents are close to each other at the top and still far from finishing long professional tasks on their own. Claude Opus 5.5 leads, Gemini 4 Carbon looks like the value pick once it ships, and every model still needs a human checking its work. For background on how these systems are trained, see our guide to RLHF.

About Teck Hustlers Editorial Team

The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.