Six frontier AI models fall short of a human-employee bar in new workplace test
Epoch AI released its Epoch Automation Reports on October 8, 2026, testing six leading large language models on tasks taken from its own day-to-day work. None reached the quality expected of a human Epoch employee. Human graders scored the outputs against Epoch's internal standards, with 100% representing an employee-level result.
Claude Fable 5.1 and GPT-6 Astra tied for first, each averaging about 65%. Gemini 3.8 Flash came last at 42%, more than 20 percentage points behind the leaders. The tasks were open-ended and included graphic design and research design. Models did relatively better on structured work such as coding and data analysis, and all of them struggled where the job called for judgment or tailoring to a specific audience. Rankings were unchanged from October 8 through November 17, 2026.
This matters for knowledge access because the promise of cheap machine intelligence rests on how much expert work it can actually replace. Standard benchmarks can flatter models. A test built from real tasks, graded against a professional bar, gives a more grounded picture. The pattern is useful: structured, checkable work is nearer to being cheap and abundant, while work that depends on judgment still needs skilled people.
The caveats are real. Only six models were tested, and just three are named in the reporting. The 100% mark is Epoch's own standard, applied by its own graders, and the tasks come from one organization's work, so results may not carry over to other workplaces. No per-task breakdown or sample sizes were given. Watch for new model releases that move the rankings, and for outside groups replicating this approach.
Source: Google News
Many Minded