3 November 2026
New benchmark finds top AI agents complete only 12% of real tasks
First reported
Latent Space ran this on .
- A new test called GDPx ran 206 multi-step tasks across economically valuable fields such as software, finance and logistics.
- The best-performing AI agent, the kind that carries out tasks on its own, completed about 12% of them.
- One model, Gemini 4 Carbon from Google, was highlighted for delivering results at lower cost than the rest.
Our read
Only 1 of 30 newsletters covered this benchmark, and Latent Space was the sole outlet to report it on 2026-11-03. It follows the August 19 study on AI agents and open-ended research tasks, and the October 19 Epoch AI grading of six frontier models on real research tasks.
Earlier on this story
Written by Newskeryx from which newsletters covered this story and from our own archive.
How it was covered
Latent Spaceswyx & Alessio
Benchmark design and domains, model rankings, cost comparisons, safeguard effects, possible bugs, audit and open-source criticism