
Illustration
Developments2 min read
Epoch test: Fable 5.2 beats GPT-6.2 Astra by 17 points on real work
Claude Fable 5.2 clearly outperformed GPT-6.2 Astra in a real-work test by research institute Epoch AI. Astra averaged 65%, 17 points lower. A second finding of the November 17 study matters more: with one exception, every model rated its own work above the better human reference solution.
Key facts
- On November 17, 2026, Epoch AI tested Claude Fable 5.2, GPT-6.2 Astra and four other models on four real tasks from its own work; no model met its staff standard on any task.
- GPT-6.2 Astra averaged 65%, 17 points below Fable 5.2; on data center research Fable 5.2 scored 97.5%, Claude Opus 5.5 87.5% and Astra 65%.
- Every model rated its own outputs above the human gold standards, with one exception: Gemini 4 Argon on the hidden-messages task.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
Epoch AI tested Claude Fable 5.2, GPT-6.2 Astra and four other models on work from its own operations: researching an AI data center, designing an experiment with chess variants, hiding messages in SVG images and writing a short data piece with a chart. Epoch staff graded the results by hand. A score of 100% means Epoch could publish the result without changes.
The results at a glance
- Overall: GPT-6.2 Astra averaged 65%, 17 points below Fable 5.2, even though the two are roughly level on the Epoch Capabilities Index.
- Data center research: Fable 5.2 scored 97.5%, Claude Opus 5.5 87.5% and Astra 65%. An outside contractor needs about 16 hours per data center, according to Epoch. Opus 5.5 took just under three hours at lower cost; Fable 5.2 took about 6.5 hours at higher cost than the contractor, but with better quality.
- No model met the standard Epoch expects from its own staff on any task.
Compared with GPT-6 Astra, Astra improved at concise writing and charts, Epoch said. Its typical mistake was overcaution: it covered only one building although the sources pointed to several, and skipped estimates it could have made.
Epoch also had the models compare their own outputs with the human reference solutions. All of them rated their own work above the better human work on all four tasks, with one exception: Gemini 4 Argon on the hidden messages. The models justified this with surface thoroughness, such as the number of experiments. Epoch itself notes the limits: a small task set, one run per model and task, and many subjective criteria. No marketing tasks were included.
Analysis
Given the small sample, the overall winner is the weaker result. The self-assessment finding is more robust, because it holds across every model. It means that asking the same model, or the same chat, to review its draft is not an approval process. Teams that sign off campaign copy, competitor analyses or decks that way are measuring effort, not quality.
Anthropic's Claude Code team describes another approach in a November 16 blog post: a separate reviewer that grades against a checklist, which raised success rates by up to ten points in internal tests. The review rules are read from the main branch so a change cannot rewrite its own checks. Marketing teams can adopt the same idea: a different model or a person reviews against a fixed checklist, and every recurring task has a reference solution that new results are measured against. For research with many sources, Epoch shows a clear gap between models. On budget, see Microsoft's MAI Code 2 and our guide to AI agents for marketing.
Got AI news? Submit news →
FAQ
Is Claude Fable 5.2 better than GPT-6.2 Astra?
On Epoch AI's real-work test from November 17, 2026, yes: GPT-6.2 Astra averaged 65%, 17 points below Fable 5.2. Epoch calls the scores noisy because the sample is small.
Can AI reliably review its own work?
According to Epoch AI, no: every tested model rated its own outputs above better human reference solutions, with one exception for Gemini 4 Argon.
Which tasks did Epoch AI test?
Four tasks from its own work: researching an AI data center, designing an experiment with chess variants, hiding messages in SVG images, and a short data piece with a chart.
Source
Epoch AI: How good are Fable 5.2 and GPT-6.2 Astra at real work?
More sources
- Claude Code Blog – 16.11.2026 (Addy Osmani), interne Testangabe von Anthropic zum getrennten Prüfer (+ bis zu 10 Punkte).



