The top AI model gets only 12% on the new GDPx test

SemiAnalysis made GDPx, a new test of AI agents (software that does tasks). It tests cyber tasks, quant trading, software engineering and 3 more domains. The top model gets 12%. SemiAnalysis says that GDPval, a test that OpenAI made in 2025, does not fit agents of today. It gives the model all the facts in the task text.

How it works

SemiAnalysis expects that the first trillion dollars of AI revenue will come from some domains. It does not include government, real estate and insurance. SemiAnalysis says that labs pay $5k or more for a good task.

The numbers

The safety filters of Opus 5.5 and GPT-6 Astra can stop answers in cyber, biology and alignment. Results by domain:

• Opus 5.5: top overall, and first in quant trading, hardware hardware design and chip design.
• Gemini 4 Carbon: first in biology, and first by a small margin in software engineering.
• GPT-6 Astra: first in AI self-improvement and in materials science.

Muse Wonder Full is the top cyber model. In finance and legal tasks, all models get 0.

What is not known

The text gives no result for each model. SemiAnalysis says that its results agree with its view of the models. Gemini 4 Carbon is not in general use, and users cannot test it.

Insight: Safety settings can change which model is first

SemiAnalysis says that safety filters lower the results of Opus 5.5 and GPT-6 Astra. A Cyber Index report shows that GPT-6 Sol with low cyber guardrails got 32 points more than the version for all persons. Only members of one OpenAI program get it. This reads as a change in access: the top model changes with the safety level that you get. Tell your vendor which level you must have.