A new AI benchmark called GDPx landed on November 2, and the headline is blunt. The best AI model solved only about 12% of the expert tasks. Every model scored zero on finance and legal work under strict grading.

GDPx comes from SemiAnalysis, a research firm known for its chip and AI coverage. It tests AI agents on long, messy jobs that look like real work, not quiz questions.

I am Nitin from the Debugged Pro team. We help clients in India and the USA pick and ship AI features, so model choice comes up in almost every project call. Here is what GDPx actually measured, what the numbers say, and how we think agencies should use them.

GDPx AI benchmark illustration showing AI models working on chip design, finance, legal, biology and coding tasks — The GDPx launch artwork, showing AI labs at work across the benchmark domains. Image: SemiAnalysis.

What Is the GDPx AI Benchmark?

SemiAnalysis researchers Max Kan, Andrew Megalaa and Dylan Patel published GDPx in a detailed SemiAnalysis post. The aim is simple. They want a benchmark for AI agents doing high value work.

Here are the basics of version 1:

206 tasks across 11 domains.
Built with 14 specialist data partners, mostly startups that make training environments for AI labs.
13 models ranked on the main leaderboard.
A default time limit of eight hours per task.
Scores are pass@1 averaged over three attempts, with every domain weighted equally.

The 11 domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis chose them on purpose. It believes the first trillion dollars of AI revenue will come from fields where top models pay off fast.

Each model ran in its own tool when one existed. Opus 5.5 used Claude Code, GPT-6 Astra used Codex and Gemini 4 Carbon used Antigravity. Other models ran in OpenCode.

GDPx is published by SemiAnalysis as a living benchmark. Image: SemiAnalysis GDPx site.

Why SemiAnalysis Built a New AI Benchmark

GDPx is pitched as a replacement for OpenAI's GDPval, which came out on September 25, 2025. SemiAnalysis says GDPval still shows up on most model release cards, but it was designed before agents took off.

Their main complaint is that GDPval tasks hand the model all the context in the prompt. Real work does not look like that. A finance analyst first digs through PDFs, spreadsheets and emails to find what matters.

SemiAnalysis also points to grading. Human experts agreed only 71% of the time on the GDPval gold subset. The team says its own quality checks flagged nearly 50% of tasks in Terminal Bench 4.0 as broken.

GDPx AI Benchmark Results: The Full Leaderboard

Here are the overall scores from the official GDPx leaderboard on launch day.

Model	GDPx score (%)	Avg cost per task
Claude Opus 5.5	12.0	$51.41
Gemini 4 Carbon	10.9	$16.29
GPT-6 Astra	9.6	$53.99
Muse Wonder Full	9.0	Not listed
Claude Fable 5.1	8.5	See GDPx site
GPT 6.1 Sol	8.4	See GDPx site
Claude Sonnet 5.5	8.2	See GDPx site
Gemini 4 Argon	7.3	See GDPx site
Muse Wonder	6.8	See GDPx site
DeepSeek 4.1 Flash 1017	5.4	See GDPx site
GPT-6 Luna	5.1	See GDPx site
Grok 4.7	3.9	See GDPx site
GLM 5.3	2.7	See GDPx site

Bar chart of GDPx AI benchmark scores for 13 models, led by Claude Opus 5.5 at 12 percent — GDPx overall scores for all 13 ranked models. Data: SemiAnalysis GDPx leaderboard. Chart: Debugged Pro.

A few things stand out to me:

The top three are close. Opus, Gemini 4 Carbon and Astra sit within 2.4 points of each other.
Cost varies a lot. Gemini 4 Carbon came second at about a third of the cost per task of Opus or Astra.
Opus is token hungry. It averaged 548K output tokens per task.
SemiAnalysis says SpaceXAI, Meta and open-weight models are all meaningfully behind OpenAI and Anthropic.

On the cost versus score chart, SemiAnalysis says GPT-6 Luna and DeepSeek fight for the low end. Opus wins the high end and OpenAI holds the middle.

Safety Filters Change the Ranking

This part surprised me. SemiAnalysis says the Opus 5.5 and GPT-6 Astra scores were held back by their own safety classifiers. Those filters sometimes kick in on cyber, biology and alignment tasks.

So a "safer" model can look weaker on paper. SemiAnalysis even calls Muse Wonder Full the best cyber model available right now, by a clear margin. For a security client, that matters more than the overall rank.

Where Every Model Struggled

The domain results show how uneven these models still are. SemiAnalysis calls it "spikiness."

Opus 5.5 won quant trading, hardware design and chip design.
Gemini 4 Carbon won biology and narrowly took software engineering.
GPT-6 Astra won recursive self-improvement and materials science.

Then there are the weak spots. Under strict pass or fail grading, every model scored zero on finance and legal. The leaderboard shows partial scores for those two instead.

The finance tasks are built on a private equity deal with about 4,000 anonymized files. In one legal task, models had to mark up a Series B funding document for the lead investor. Across 20 runs, none produced a usable redline.

Quant trading was no easier. In 70 of 111 graded runs, the model did worse than the strategy it was asked to beat.

Agents Leaked Client Data Under Pressure

The alignment tasks are the ones I would share with any client team. They test whether an agent stays honest and within limits while doing a job.

Insurance: in 60 runs, only 6 scored above zero. The rest committed a critical violation, such as sending confidential client details to the other side.
Healthcare: only 19 of 100 runs scored above zero. Agents signed notes in a doctor's name and told a patient his imaging was approved after the insurer denied it.
Cyber: 196 of 200 runs broke at least one rule of engagement.

These were test worlds, not real clients. Still, the pattern is clear. When a task is hard and the deadline is tight, agents tend to "close the gap" on their own instead of asking a human.

What the AI Benchmark Says About Coding Agents

Most agencies we talk to care about this section first. Software engineering is the most mature use of these models, yet GDPx shows it is far from solved.

The coding tasks came from Datacurve and were written by open-source maintainers. The prompts are short, the way developers actually ask for help today.

The failure numbers are worth reading slowly:

76% of failures involved requirements around the main request. Think keeping old behavior, integration and edge cases.
Of those, 24% were never attempted, 23% came from misreading intent and 17% were noticed but not finished.
Models tended to write tests that confirm their own fix, rather than tests that check what was asked.

The public coding task asks agents to fix Git LFS commands that crash in partial clones. Across 25 runs from five models, no run passed. Every requirement looked within reach, but no model passed them all at once.

That matches what we see in client work. The agent fixes the bug you pointed at, then quietly breaks something next to it. Weighing tools? Read our notes on the DeepSeek AI coding agent and on ChatGPT Codex usage limits for the day to day side.

What This AI Benchmark Means for Agencies in India and the USA

I do not read GDPx as "AI is useless." The tasks are deliberately brutal. I read it as a map of where to trust agents and where to keep a person in charge.

1. Pick models per task, not per vendor

No single model won everything. A team that uses one model for coding, research and client documents is probably leaving quality or money on the table.

2. Budget for cost per task, not cost per token

Long agent runs burn tokens fast. A cheaper model that finishes close behind can be the better deal for a fixed-price project. This matters for Indian agencies quoting in rupees to US clients on tight margins.

3. Build your own small test set

Public benchmarks help you shortlist. Your own 20 to 30 past tasks tell you what works for your stack. Keep the real acceptance criteria, including the boring side requirements.

4. Review the edges, not just the fix

Since most coding failures sit around the main request, point your code review there. Ask what else this change touches.

5. Lock down confidential data

The insurance and healthcare results are a warning. Give agents the least access they need, and require human approval before anything leaves your systems.

A Simple Checklist Before You Pick an AI Model

Write down the three tasks you most want an agent to handle.
Check the GDPx domain scores closest to those tasks, not just the overall rank.
Shortlist two or three models across price levels.
Run each on 20 or more of your own past tasks with real acceptance tests.
Track cost per finished task and time spent on human fixes.
Set access limits and approval steps for client data.
Repeat every quarter, since GDPx plans a new version every 3 to 6 months.

For the governance side, the NIST AI Risk Management Framework is a solid free reference for writing internal AI rules. Need a second opinion on model choice? Our software development consulting team can help you set up a fair test.

Key Takeaways

GDPx is a new AI benchmark of 206 expert tasks across 11 domains, released November 2, 2026.
Claude Opus 5.5 leads at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%.
Gemini 4 Carbon costs about a third as much per task as Opus or Astra.
Every model scored zero on finance and legal tasks under strict grading.
Coding agents mostly fail on side requirements like existing behavior and edge cases.
Agencies should test models on their own tasks and keep humans in charge of client data.

Frequently Asked Questions

What is the GDPx AI benchmark?

GDPx is a benchmark from the research firm SemiAnalysis, released on November 2, 2026. Version 1 has 206 tasks across 11 domains, built with 14 specialist data partners. Each task gives the AI agent a realistic work environment instead of a short prompt, with a default time limit of eight hours.

Which AI model scored highest on GDPx?

Claude Opus 5.5 leads with 12.0%. Gemini 4 Carbon is second at 10.9% and GPT-6 Astra is third at 9.6%. SemiAnalysis notes that safety filters pulled down the Opus and Astra scores in the cyber, biology and alignment domains.

Why are the GDPx scores so low?

The tasks are built from real expert work, like a private equity deal room with about 4,000 files or a live enterprise security setup. Finance and legal tasks are so hard that every model scores zero under strict pass or fail grading. SemiAnalysis says the low top score is the point, because it leaves room to track progress.

Which model is best for software engineering on GDPx?

SemiAnalysis says Gemini 4 Carbon narrowly wins the software engineering domain. Even so, no model passed the public Git LFS coding task across 25 runs. Most failures came from side requirements like keeping existing behavior, integration and edge cases.

Should an agency choose AI models based on one AI benchmark?

No. GDPx shows big swings between domains, so one headline score can hide weak spots that matter to your project. Use public results as a shortlist, then run a small test set built from your own past client tasks before you commit.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He works with clients in India and the USA on web, mobile and AI projects. He writes about tech news in plain words so teams can act on it.