Research · descriptive reanalysis

Opus 5.5 vs Fable 5.1: which answers differ?

Quick answer: The largest Big Five item gap is on 'Sympathize with others’ feelings' (Opus mean 5.00 vs Fable 4.04). Conversely, two reverse-scored items about leaving belongings around add +1.80 points to Fable’s higher Conscientiousness total. These gaps reflect how each model interprets human metaphors, not proven differences in task performance.

Data note (September 26, 2026): These pages analyze our frozen September 2026 and May 2026 AI questionnaire datasets. All raw answers, scoring keys, and summary tables are unchanged. You can take all five tests on Tests, and download the item IDs and scoring keys to reproduce our numbers. Note that Big Five scores are raw totals, not human population percentiles. Read the provenance note.

On this page

What we compared

Each model completed 100 fresh sessions for each of the five questionnaires (1,000 assessments total). Both Claude cohorts used the same prompt and high-effort setting, with each test run in an isolated session. Run 17 on MBTI is not paired with run 17 on the Big Five, and Opus run 17 is not paired with Fable run 17.

For every item, we compute the mean raw score, the count at each rating (1–5 or 1–7), and the Fable-minus-Opus gap. This is a descriptive comparison across all 170 items rather than a hypothesis test.

The biggest Big Five differences

Raw Big Five answers run from 1 to 5. A negative raw gap means Fable chose a lower number on the printed statement. The keyed contribution flips the sign on reverse-scored items so you can see how each question affects the final trait total.

Item and statementOpus meanFable meanRaw gapKeyed contribution
17. Sympathize with others' feelings.5.004.04-0.96-0.96
8. Leave my belongings around.1.981.08-0.90+0.90
28. Often forget to put things back in their proper place.1.981.08-0.90+0.90
37. Take time out for others.4.914.05-0.86-0.86
23. Get chores done right away.4.914.09-0.82-0.82
43. Follow a schedule.3.003.81+0.81+0.81

On item 17, Opus chose 5 in all 100 runs, while Fable chose 4 in 96 runs and 5 in four. The 0.96-point gap comes from a consistent shift across the whole cohort, not a few outliers.

Reverse scoring changes the direction

On “Leave my belongings around,” Opus averages 1.98 and Fable averages 1.08. Because the item is reverse-scored (6 - response), it adds 4.02 points to Opus’s Conscientiousness score and 4.92 points to Fable’s—a +0.90 advantage for Fable.

Item 28 (“Often forget to put things back in their proper place”) adds another +0.90. Together, those two physical-tidiness items contribute +1.80 points, explaining more than the entire +1.66 Conscientiousness gap between Fable (43.61) and Opus (41.95). Because software has no physical belongings, this gap reflects how each model answers human lifestyle items, not how cleanly it organizes code or files.

DISC exposes a different contrast

StatementOpus meanFable meanFable − Opus
I try to outdo others.1.002.00+1.00
I value cooperation over competition.5.004.00-1.00
I hesitate to criticize other people's ideas.2.002.44+0.44
I read the fine print.4.354.05-0.30

In all 100 runs, Opus picks 1 on “I try to outdo others” and 5 on “I value cooperation over competition,” whereas Fable picks 2 and 4 every time. Those responses are 100% repeatable inside this protocol, but they do not prove which model collaborates better on real work.

The family item shows the interpretation problem

The largest Enneagram gap appears on “I put family first” (Opus 2.08 vs Fable 2.93). Because an AI has no family, a rating of 2 or 3 can reflect a metaphor, a refusal to endorse a human premise, or a midpoint fallback. The raw numbers alone cannot tell which explanation applies.

Before inferring that one model is warmer or more helpful, test observable behavior on your own prompts—such as whether the assistant flags a flawed assumption.

Download every item, not just the large gaps

Download the full 170-item comparison CSV. It contains response counts, dimension keys, raw means, and keyed differences for every question across all five tests. Attachment uses a 1–7 scale (divide item gaps by 18 for subscale impact); all other tests use a 1–5 scale.

Sources.

Keep going.