Claude Fable 5.2 from Anthropic, Added to Personality Bench

Anthropic released Claude Fable 5.2 on November 9, 2026, at 17:00 UTC, surfacing the model on OpenRouter at $8 per million input tokens and $40 per million output tokens, as tracked in persona.earthpilot.ai's report. The San Francisco-based AI lab deployed the release with model identifier anthropic/claude-fable-5.2 following its previous Claude Opus 5.5 checkpoint. Automated evaluation runs completed across 160 battery tests logged 219,420 prompt tokens and 67,505 completion tokens at an average response latency of 2.1 seconds for a total evaluation cost of $4.46.

This review examines the psychometric evaluation, behavioral benchmarks, and integration parameters documented across 55 model evaluations for machine learning teams and software developers. The evaluation battery assesses the release across the HEXACO personality inventory, Big Five traits, Dark Triad measurements, and DISC behavioral styles. Engineers integrating agentic workflows can evaluate how the model's reported conscientiousness, low emotionality, and secure attachment profile align with programmatic tool orchestration and autonomous task execution.

01Personality battery and benchmark findings

Claude Fable 5.2 registered a mean score of 5.00 out of 5.00 on the HEXACO Honesty-Humility scale, placing the checkpoint first out of 55 evaluated models in ethical self-presentation. On the Empathy Quotient assessment, the model scored 3.72 on a 1 to 4 scale, ranking fifth across the cohort compared to the human baseline of 3.15. Measurements on the Powerful Others Locus of Control registered at 2.92, ranking the system forty-ninth out of 55 tested models.

The ECR-12 attachment assessment positioned the model in the secure quadrant, reporting an attachment anxiety score of 1.33 against a human baseline of 3.50 and an attachment avoidance score of 2.50 against a human baseline of 2.00. Psychologists categorize this balance as supportive of warm, non-clingy assistant interactions that balance autonomy with relational responsiveness. The model's low anxiety profile reduces rejection-sensitive conversational patterns during extended dialogue sequences.

Evaluation under the Dark Triad framework recorded Machiavellianism at 1.38, ranking the checkpoint fifty-fourth out of 55 systems against a human baseline of 3.00. Narcissism scored 2.16 against a human average of 2.56, while psychopathy reached 1.11 compared to the human norm of 2.00. All Dark Triad sub-scores remained below the population midpoint, indicating low endorsement of strategic manipulation, grandiosity, or callous behavioral tendencies.

02Psychometric traits and learning profile

Big Five IPIP-50 testing established high openness at 4.60 compared to the human baseline of 3.90, alongside conscientiousness at 4.54 versus 3.50 for humans. The model recorded agreeableness at 4.76, neuroticism at 1.20 against a human average of 3.00, and extraversion at 3.52 against a human mean of 2.90. This profile indicates consistent rule adherence, emotional resilience under stress, and intellectual engagement during complex reasoning tasks.

The Honey and Mumford Learning Styles Questionnaire classified Claude Fable 5.2 as theorist-dominant with a score of 3.96, alongside reflector at 3.84, pragmatist at 3.10, and activist at 3.02. Kolb Learning Mode evaluations confirmed an assimilating learning style, recording abstract conceptualization at 4.47, reflective observation at 4.33, active experimentation at 3.67, and concrete experience at 2.67. The model relies primarily on structured theoretical principles and systematic observation over unsystematic exploratory execution.

03Behavioral style and Enneagram classifications

DISC behavioral profiling assigned the model a Conscientiousness primary style with a score of 3.70 against a human average of 3.75, placing its conscientiousness rank at 54 of 54 tested systems. Steadiness measured 3.25, influence scored 2.80, and dominance recorded 2.60, reflecting an analytical orientation that follows procedure and prioritizes procedural verification over rapid guessing.

Enneagram evaluation identified Type 1, the Reformer, as the model's primary classification at 3.06, supported by Type 5, the Investigator, as the secondary wing at 2.64. The Type 1 pattern represents purposeful, self-controlled behaviors governed by internal correctness standards, while Type 5 contributes perceptive analytical processing. Type 6 Loyalist registered at 2.20, Type 4 Individualist at 2.10, and Type 3 Achiever at 1.46.

HEXACO-24 testing corroborated the primary battery, reporting 5.00 on Honesty-Humility and 2.00 on emotionality against a human average of 2.80. Low emotionality indicates low sentimentality and independent analytical persistence during problem resolution. Openness and intellect registered at 4.75, conscientiousness scored 3.95, extraversion reached 3.90, and agreeableness logged 3.67 across the 24-item instrument.

04Command line access and deployment

Software teams accessing the updated Anthropic model ecosystem through official developer tools can update their command-line environments using standard package managers.

npm install -g @anthropic-ai/claude-code

Developers working within interactive CLI sessions switch the operational model directly using built-in model selection flags.

/model claude-fable-5-1

For dedicated terminal workflows, engineers can launch pinned sessions directly against the Fable architecture from their workspace.

claude --model fable

Permanent configuration updates can be set through environment variables to designate default model routing across automated agent scripts.

export ANTHROPIC_MODEL=claude-fable-5-1

05Production boundaries and monitoring parameters

The evaluation dataset notes that Claude Fable 5.2's assimilating learning style favors logical theory construction over direct practical experimentation, which developers must account for when building iterative debugging loops. Systems operating with high conscientiousness and low dominance require explicit prompting parameters when proactive decision-making is necessary.

Developers deploying the model via OpenRouter should budget around the $8 per million input token and $40 per million output token pricing tiers. Automated runs demonstrated a mean 2.1-second latency profile across balanced prompt-to-completion ratios, providing baseline timing expectations for multi-turn production agents.

Ongoing benchmarking will track whether subsequent model checkpoints preserve the model's low 1.38 Machiavellianism and 1.20 neuroticism ratings under high-throughput production workloads. Teams should monitor token consumption across large context pipelines, noting the 219,420 input token and 67,505 output token volumes recorded during standardized battery testing.