No AI model met Epoch AI’s own employee standard on real Epoch work tasks: Claude Fable 5.2 led at 82% and GPT-6.2 Astra averaged 65%, ranking fourth of six, on the research group’s updated Epoch Automation Reports benchmark published on 17 November 2026.
Our number: Astra finished 17 points behind Fable 5.2 even though Epoch says the two share the top Epoch Capabilities Index score of 174; the leader was still 18 points short of the 100% employee bar (our arithmetic).
Business Pill · Least Privilege
An evergreen Business Engineer explainer on access control for software agents. It teaches the general idea only and does not discuss this benchmark.
The key insight: On Epoch AI’s own work tasks, no model reached the employee standard: Claude Fable 5.2 led at 82% and GPT-6.2 Astra averaged 65%, 17 points lower despite the tied top capabilities index score of 174 that Epoch reports.
What Epoch Measured
Epoch describes the benchmark as “A benchmark of realistic, open-ended tasks drawn from Epoch AI’s own work”. “A score of 100% means the output meets the internal standard we’d expect from an Epoch employee”, and “A human grader then scores each output against rubrics that reflect our internal quality standards”.
Tasks fall into four categories: graphic design, data insight generation, AI data center research and research. Epoch says it runs “each model on its corresponding agent harness at its highest reasoning setting” and gives models “as much context as we would give a human contractor”.

What Was Posted on X
Epoch AI posted on 17 November 2026 that no model it tested met its employee standard on Epoch Automation Reports, with Fable 5.2 highest at an 82% average and GPT-6.2 Astra fourth at 65% despite sharing the top ECI score of 174 with Fable 5.2.
How the Six Models Scored
Epoch’s average scores: Claude Fable 5.2 82%, Claude Opus 5.5 78%, Gemini 4 Argon 67%, GPT-6.2 Astra 65%, Claude Fable 5.1 64% and GPT-6 Astra 56%. All six averaged below 100%.
The widest gap was in AI data center research, where Fable 5.2 scored 98% and GPT-6.2 Astra 65%, a 33-point difference (our arithmetic). GPT-6 Astra scored 40% on that category.
On that task, Epoch wrote on X: “Fable 5.2 and Opus 5.5 performed at the level of a human contractor on our AI data center task, which involves researching EcoDataCenter 1”. It said GPT-6.2 Astra “was overly cautious and left out critical information”.
Astra Improved, but Not Enough to Close the Gap
GPT-6.2 Astra averaged 65% against 56% for GPT-6 Astra, a 9-point gain (our arithmetic). Epoch’s earlier post said Astra “improved only slightly over GPT-6 Astra” and that “Its writing was relatively strong compared to the other models we evaluated, but still below Epoch AI’s standards”.
The result echoes our report on testers panning GPT-6.2 Astra and sits beside Epoch’s earlier capabilities index readout. It also lands the same day as a Google paper on agent harness memory and after OpenAI’s first ‘High’ self-improvement rating.
The Structural Read
As we read Epoch’s results, a tied general capabilities score did not predict performance on open-ended work graded by the people who do that work.
As we read it, the 98% Fable 5.2 scored on AI data center research shows one task category close to contractor level, while the averages stay well below the employee bar.
Epoch AI, on X
“All models failed to meet the Epoch employee standard.”
Three Implications
FOR MODEL BUYERS Epoch reports a 17-point gap between two models that share the top capabilities index score of 174.
FOR AUTOMATION PLANS Epoch says Fable 5.2 and Opus 5.5 performed at the level of a human contractor on its AI data center task.
FOR READING BENCHMARKS Epoch notes it ran each model once per task and that harness differences may affect results.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The analysis puts it this way: “Foundation models become relatively commoditized as orchestration layers capture differentiation.”
As we read Epoch’s note that some differences may come from the harness rather than the model, results like these test the model and its harness together. The analysis does not discuss Epoch’s benchmark.
What Is Not Established
Epoch lists its own limits: “We have a limited task suite, and we run each model once per task”, and “These results should not be read as a general measure of what AI can or cannot do across all work tasks”. It also notes that “Some differences we observe may be influenced by the harness rather than the model itself”. Epoch says a full report on the updated benchmark will follow.
The Bottom Line
On Epoch’s own work, the best model reached 82% of the employee standard and GPT-6.2 Astra 65%, 17 points apart despite a tied capabilities index score.
97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. All scores, method details and quotations are from Epoch AI’s Epoch Automation Reports page (17 November 2026) and Epoch AI’s own posts on X, read by us. The 17-point, 18-point, 33-point and 9-point gaps are our arithmetic from Epoch’s published averages. We did not contact Epoch AI, OpenAI, Anthropic or Google.
Sources: Epoch AI, Epoch Automation Reports (17 Nov 2026) · Epoch AI on X, 17 Nov 2026 · Epoch AI on X, 17 Nov 2026 (earlier post)









