What happened
On October 8, 2026, Epoch AI published Automation Reports, a set of open-ended tasks taken from its own work: graphic design, data insight, a data explorer, AI data-center research, and research design. A person grades each output against the standard Epoch expects of an employee. Each model was run once per task, on a high reasoning setting. Claude Fable 5.1, in Claude Code, and GPT-6 Astra, in Codex, both average 65 percent. Grok 4.6 averages 59 percent, Qwen 3.8 Max 53 percent, Kimi K3 52 percent, and Gemini 3.8 Flash 42 percent. Epoch says the lead is mostly reliability on the defined parts, such as coding and calculation, and that the open-ended parts still fail. Epoch removed one graphic task, restyling a chart to house style, because Fable 5.1 produced an acceptable chart, and now partly automates that step.
Why it matters
A benchmark that a model can saturate is a different object from a job a colleague would accept. Epoch's point is that 65 percent of the rubric items is still a rejection: a person has to check and fix the work. The organization also records the other direction. One narrow chart task has already crossed the line, and they took it out of the suite. The useful reading is task by task, not a single percent of the economy.
Fable's category scores are 65 percent on graphic design, 70 on data insight, 66 on the data explorer, 63 on data-center research, and 63 on research design. Astra's are 73, 77, 46, 67, and 62. The averages match. The category rows do not.
The harnesses were Codex for Astra at Ultra, Claude Code for Fable at Ultracode, Grok Build for Grok 4.6 at xhigh, Antigravity for Gemini 3.8 Flash at High, Kimi Code for Kimi K3 at Max, and Qwen Code for Qwen 3.8 Max at Max.
Epoch says the sample is small, each model ran once, and a different grader or a different rubric could move the numbers. The tasks are Epoch's, so they are not a sample of other offices.