What five AI models say about themselves.

Abstract

We reanalyzed five historical AgentTune questionnaire adaptations administered on September 24, 2026: an MBTI-style instrument (OEJTS), Big Five (IPIP-50), DISC (ODAT), Enneagram (OEPS) and attachment (ECR-R). We recovered and validated 2,000 response vectors for Astra, Sol, Opus and Fable, plus three canonical Grok vectors. The five scales contain 170 items per full battery. We report raw means, original labels, exact ties and protocol differences.

Astra and Sol differ in their MBTI distributions, but different prompts prevent a model-only explanation. Opus reports higher agreeableness than Fable under a closely matched questionnaire protocol. Exact ties affect 29% of Astra MBTI runs and 51% of Fable DISC runs. These are exploratory model self-descriptions under specific prompts, not validated psychological traits, observed behavior, a performance ranking or evidence that tuning improves outcomes.

Five models, three kinds of evidence

For Astra and Sol, five assessments share one session; they are not 500 independent model invocations. The Claude instruments were administered separately, so run 001 on MBTI is not a paired person or session with run 001 on DISC. Fresh sessions reduce conversation carryover but do not establish statistical independence. No tuning was installed.

Opus and Fable saw the same questions, scales and neutral system prompt; test names, keys and prior results were withheld. Their CLI builds differ. Astra and Sol used named, fixed-order batteries inside Codex, with different instructions for adapting human experiences to assistant behavior. Default scaffolding remained. Sampling temperature and random seed were not controlled. Three supplied Astra reports contain the same 100 batteries; we counted that cohort once.

MBTI: keep the original result, reveal the tie

The original scorer sent exact ties to I, N, T and J. Removing that default leaves Sol with 79 fully resolved INTJ runs rather than 97. Opus has 84 resolved INTJ runs rather than 94; Fable has 98 rather than 99. Astra has 45 resolved ISTJ and 19 resolved INTJ runs. The other results are either a different resolved type or a pattern with X for a tied axis. The downloadable summary contains every pattern.

This is a scoring sensitivity result, not a new round of model answers. Astra’s split should not be described as a switch from an established INTJ personality. Grok’s supplied canonical vector scores INTJ with no tied axes, but its simulated INTJ rate is not a replication rate from independent sessions.

Big Five: compare raw scores before human percentiles

Each trait sums ten items, with a possible range of 10–50. These are means over 100 responses per instrument, except Grok’s reported canonical profile, marked with an asterisk.

* Grok’s Big Five item vector was not supplied, so that row is reported rather than independently recomputed. Across the four raw-data cohorts, openness is 43.93–46.87, conscientiousness 41.95–43.73 and agreeableness 44.99–47.07. These common responses support a narrower description than “every AI has the same personality.” They may also reflect assistant instruction and item interpretation.

The closest comparison is Opus versus Fable: Opus’s agreeableness mean is 2.04 points higher, its conscientiousness 1.66 lower, and its neuroticism 1.72 lower. The sample standard deviations for agreeableness are 0.46 and 0.44. These describe this collection, not uncertainty over all possible prompts or future model versions. We do not interpret a human norm conversion as an AI’s percentile in a human population.

DISC and Enneagram: order can masquerade as preference

Counts are out of 100 per instrument. These columns describe different tests and should not be added as a count of unique sessions.

Fable’s original DISC labels are SC in 91 runs and CS in nine. Yet Steadiness wins outright only 40 times; Conscientiousness wins nine; 51 are tied. The old stable sort put S ahead of C on equality. Opus has 64 outright S wins, four C wins and 32 ties. Sol has 53 C wins, eight S wins and 39 ties. Astra has 99 S wins and one C win, with no ties.

Enneagram has the same issue with numerical order. Opus is labeled Type 2 in 54 runs under the old rule, although Type 2 wins outright only twice. Sol’s 88 Type 5 labels include 40 ties with Type 8. Fable has 55 outright Type 8 wins, one Type 5 win and 44 tied results. Astra has six outright Type 1 wins, 50 Type 2 wins, four Type 5 wins and 40 ties. Report the set of tied leaders before interpreting a core type. A wing must be adjacent to a uniquely selected core type; the older “5w2” label is not a valid wing.

Grok’s canonical vectors reproduce Type 5w4 and DISC C. They do not establish that it is more candid, autonomous or accurate on real tasks. We have also updated the interactive quizzes to show equal top scores without automatically selecting a tuning.

Attachment: a shared label with different coordinates

All four raw-data cohorts fall in the Secure quadrant on every run. That does not make their response patterns identical. Astra’s avoidance mean is 2.31; Opus’s is 3.35. The item prompts explicitly or implicitly adapt human relationship language, and that adaptation differs between protocol groups. “Secure” is a scoring label here, not evidence of an emotional bond or relationship capacity.

Grok reports anxiety 1.50 and avoidance 4.06, close to the avoidance cutoff of 4. Its 63 Avoidant / 37 Secure simulation split shows sensitivity to the assumed answer variation. Without the underlying vectors and simulation code, we cannot reproduce that split or interpret it as repeated observed behavior.

Muse Spark 1.3: a supplemental report

The supplied Muse report describes 100 administrations per instrument, with sequential runs within one agent session per test. Its aggregate MBTI results are ISTJ 80, ISFJ 12 and INTJ eight; Big Five means are O 39.5, C 41.2, E 33.8, A 44.6 and N 15.6. It reports Secure attachment in all 100 runs, with anxiety 2.04 and avoidance 2.52.

No raw vectors or harness accompanied that report. We preserve these supplied figures in the protocol download as unverified aggregates, but exclude them from the reproducible dataset and fresh-session totals. A matched rerun with archived responses would make Muse a stronger sixth comparison.

What this adds to the earlier research

The earlier research collection combines fresh responses, repeated scoring and simulations. Its roughly 2,200 reported scoring records are retained as historical reports, not pooled with this cohort. The Opus 4.8 follow-up sampled from one elicited probability profile; it is not a matched predecessor to this fresh-session study.

This release adds auditable response vectors, a frozen instrument snapshot and explicit tie counts. It also corrects the Big Five page’s prose: its item key reverses 18 items, not 20. Using that same 18-item key reproduces the submitted Big Five means. The new contribution is traceability and scoring sensitivity, not a claim that a newer model has become a different kind of person.

How to use the findings

Use the Opus guide and Fable guide as starting points for communication preferences. A high agreeableness score can motivate a test of a direct-feedback instruction; it cannot demonstrate sycophancy or prove that the instruction fixes it. Pick a few representative tasks, compare the default with one short instruction, and assess factual accuracy as well as tone.

The custom-instructions generator can produce a draft from a preferred style. These datasets contain no before/after tuning experiment. Claims about increased productivity, better decisions or user satisfaction require new outcome measurements.

Data, methods and reproduction

• Response vectors (JSON): 2,000 complete fresh-session assessments and three canonical Grok vectors; no account identifiers or raw session logs.
• Recomputed summaries (JSON) and numeric summary (CSV): counts, legacy labels, tie sets, means and sample standard deviations.
• Legacy scoring definitions (JSON): item order and numeric scoring metadata; restricted question wording has been replaced by item labels.
• Protocol and provenance (JSON): available prompts (some full questionnaire prompts withdrawn), source hashes, deduplication, missing-data notes and supplemental reported figures.
• Offline scorer (Node.js): place alongside the response and instrument JSON files and run node september-2026-score.cjs > reproduced-summary.json. It performs no model calls.

Every retained vector was checked for length, integer values and scale bounds. Claude item snapshots were identical; Astra’s frozen keys matched them. Both original labels and tie-aware summaries were recomputed from the arrays. No raw answers were edited. Scorer, data and article are versioned in the website repository. This is an exploratory research note, not a peer-reviewed or preregistered study.

Next useful experiments: randomize questionnaire order, repeat the same model under matched prompts, distinguish inapplicability from neutral responses, and test whether communication instructions change task behavior. Each requires new observations rather than more scoring of these same answers.

Questions people ask.

Sources.

• AgentTune: all five instruments and source attributions
• Earlier AgentTune research and methodology limits
• Response-level data and collection protocols

Keep going.