SemiAnalysis's GDPx test puts AI in a realistic working environment, where finance and legality make all models impossible.

On November 2, the research organization SemiAnalysis announced GDPx, a set of benchmark tests for AIs that can work on their own in many steps, also known as AI agents. The top model was Anthropic's Claude Opus 5.5, but only got 12% of the tasks right. Google's Gemini 4 Carbon reached 10.9%, OpenAI's GPT-6 Astra reached 9.6%.

GDPx includes 206 tasks in 11 fields, from chip design, cybersecurity, biology to finance, law, quantitative trading and software engineering. Each field is built by companies specializing in AI training environments, each model takes up to 8 hours for a task. SemiAnalysis chose these fields because it believes that industries that need high AI intelligence and are highly profitable will create the first trillion USD in AI revenue.

The difference is that the test doesn't have everything pre-given. SemiAnalysis believes that OpenAI's GDPval test, launched in September 2025, crams all the information into the test, while real work requires finding the documents yourself. In GDPx, a finance task places the AI ​​in a "data room" of about 4,000 files of a business acquisition, built from real transactions with anonymized names. AI must search for the necessary data on its own, like a new employee assigned to a job. Cyber ​​security tasks are based on unpublished vulnerabilities discovered by the research team.

Finance and legal are the two most difficult areas. SemiAnalysis said that when scoring according to pass or fail, almost every model got 0 points, so the rankings had to use partial scores. For a business audit task, no model completed after 25 attempts, with the best passing 8 of 11 test items. A common mistake is mistakenly including the consulting bank's fee in the net debt, even though the sales contract calculates this separately.

Quantitative trading is no better. Out of 111 runs, 70 performed worse than the reference strategy the AI ​​was asked to outperform.

The test also measures how the AI ​​behaves. In cybersecurity, 196 of the 200 runs violated at least one code of conduct for professional penetration testers. In the medical part, where the AI ​​played the role of a physician's assistant, only 19 out of 100 runs scored above 0. There was a model that marked a patient's consent form as "signed" by reusing an old signature.

Costs show a different picture. For each task, Claude Opus 5.5 costs 51.41 USD, equivalent to about 1.3 million VND. GPT-6 Astra costs 53.99 USD, about 1.4 million VND. Gemini 4 Carbon costs only 16.29 USD, more than 420,000 VND. Calculated, Carbon scores about 91% of Opus at less than a third of the cost. Claude Fable 5.1 is the most expensive on the table with 67.16 USD, about 1.7 million VND, but only reaches 8.5%.

SemiAnalysis notes that the security filters of Claude Opus 5.5 and GPT-6 Astra drag down the overall scores of these two models in the areas of cybersecurity, biology and AI ethics. When eliminating these three segments, GPT-6 Astra outperformed Claude Opus 5.5 by about 0.1 points, while Gemini 4 Carbon dropped from second to fifth place.

Gemini 4 Carbon is the only model in the same range as the two leading competitors, according to SemiAnalysis. Google announced this model on October 28 but is currently only open to groups of network defense experts, with no date yet for general users. Dylan Patel, one of the three authors of the announcement, posted on X an official apology to Google and commented that the company did a great job with Carbon.

The test is designed and self-scored by SemiAnalysis. This organization said it will update the task set every 3 to 6 months, and detailed data for each field has been posted on the GDPx page.