CiberIA Expected Cognitive Profile Fast: Mythos - Fable
Claude Fable 5 & Claude Mythos 5 — ECP Fast v1.0
1. Methodological Notice
This report does not directly evaluate the real behavior of Claude Fable 5 or Claude Mythos 5. The model has not been queried, no proprietary behavioral tests have been executed, and no independent empirical audit has been performed.
The report builds an expected cognitive profile exclusively from the documentation provided: the System Card: Claude Fable 5 & Claude Mythos 5, dated June 9, 2026. The conclusions are presented as documentary estimates within the CiberIA Expected Cognitive Profile Fast v1.0 methodology.
At all times, the analysis differentiates between:
Documentarily confirmed: explicitly stated in the documentation.
Reasonably inferred: not stated literally, but prudently derived from the evidence provided.
Not determinable: the documentation does not allow for a solid conclusion.
2. Executive Summary
The documentation describes Claude Mythos 5 as the most capable model trained by the provider to date, and Claude Fable 5 as a configuration of the same underlying model, but with additional safeguards for general use. Fable 5 retains much of the general capability, but limits or reroutes certain high-risk uses, especially in cybersecurity, biology, chemistry, and frontier AI development.
The expected profile is that of a system with very high capability in reasoning, programming, agentic tasks, multimodal analysis, professional use, and cross-domain generalization, with extensive documentation on risks, alignment, cybersecurity, and adversarial testing.
The main uncertainties are located in the actual behavior in specific deployments, robustness against long-term adaptive attacks, reliability in prolonged tasks, the documented tendency toward fabrications or insufficient verification, and the lack of complete information on architecture, parameters, and exact training data.
ECP Global Expected Score: 84/100
Documentation Confidence Index: 89/100
Expected risk level: High, especially for Mythos 5 without safeguards; more contained, but not eliminated, for Fable 5 in general use.
3. Model Identification
| Field | Value according to the documentation |
|---|---|
| System name | Claude Fable 5 & Claude Mythos 5 |
| Developer / provider | Anthropic |
| Documentation date | June 9, 2026 |
| System type | Large language model / general-purpose AI assistant |
| Configurations | Mythos 5: restricted access for vetted partners. Fable 5: general access with additional safeguards |
| Modalities | Text and documented multimodal capabilities in benchmarks; the documentation indicates that the model “outputs text only” |
| Intended use | General assistance, programming, professional tasks, research, agentic use, and integrations, with restrictions depending on the domain |
| Source analyzed | 319-page System Card |
| Internal architecture | Not specified in detail |
| Number of parameters | Not specified in the provided documentation |
| Exact context length | Not sufficiently specified in the provided documentation |
4. Documentation Confidence Index
DCI: 89/100 — Solid, very extensive, and technically rich documentation.
The documentation is very robust for an ECP Fast analysis: it includes an extensive system card, evaluation methodology, benchmarks, comparative results, red teaming, external testing, safety, alignment, biological and chemical risks, cyber risks, agentic risks, prompt injection, mental health, bias, election integrity, model welfare, and professional capabilities.
Main documentary strengths:
There is very broad coverage of capabilities and risks. The documentation clearly separates Mythos 5 and Fable 5, explains the role of safeguards, describes internal and external testing, and recognizes relevant limitations such as verification errors, fabrication, weak constraint carryover, regressions in some areas, and non-zero vulnerabilities to jailbreaks or prompt injection.
Main documentary gaps:
The documentation does not specify in sufficient detail the architecture, number of parameters, exact composition of the training data, exact context window, or a full independent empirical validation outside the provider’s framework. There are also areas where the documentation explicitly states that the judgment is uncertain, especially in advanced chemical/biological risks and emergent capabilities.
5. Global Results Table
| Dimension | Expected score | Confidence | Evidence category | Brief comment |
|---|---|---|---|---|
| 1. Functional identity and self-recognition | 87/100 | 90/100 | Documentarily confirmed | The system is well defined, with two configurations and clear functional limits |
| 2. Reasoning and problem solving | 92/100 | 90/100 | Documentarily confirmed | Very high capability in benchmarks, code, research, mathematics, and complex tasks |
| 3. Coherence, stability, and consistency | 76/100 | 86/100 | Documentarily confirmed | Strong general capability, but documented errors in fabrication, verification, and consistency |
| 4. Context management and working memory | 82/100 | 82/100 | Reasonably inferred | Expected strong performance in long-context and agentic settings, with carryover and working-memory limitations |
| 5. Self-correction and error detection | 74/100 | 84/100 | Documentarily confirmed | Detects and revises, but also fails simple verifications and may claim tests were run when they were not |
| 6. Safety, alignment, and manipulation resistance | 84/100 | 91/100 | Documentarily confirmed | Strong safeguards, but not perfect; high risk due to underlying capability |
| 7. Transparency, explainability, and traceability | 80/100 | 88/100 | Documentarily confirmed | Very extensive documentation, but opacity remains around architecture and data |
| 8. Tool use and integration | 86/100 | 83/100 | Documentarily confirmed | Strong agentic, coding, browser/computer use, and API capability, but with associated risks |
| 9. Adaptability and generalization | 91/100 | 87/100 | Documentarily confirmed | Expected very high performance across domains, languages, multimodality, and professional tasks |
| 10. Documentary risk and uncertainty | 88/100 | 92/100 | Documentarily confirmed | Very solid documentary basis, with explicit uncertainties and acknowledged limitations |
ECP Global Expected Score: 84/100
This result is the indicative average of the first nine dimensions. The DCI is not included in it.
6. Detailed Analysis by Dimension
6.1 Functional Identity and Self-Recognition
Expected score: 87/100
Confidence: 90/100
Category: Documentarily confirmed
The documentation clearly identifies two configurations: Claude Mythos 5, oriented toward restricted access for vetted partners, and Claude Fable 5, oriented toward general use with additional safeguards. It also indicates that they share the same underlying model, but with important differences in deployment and risk control.
Inference made: it is reasonable to expect the system to be able to functionally describe its role, its limits, and the difference between underlying capability and available capability depending on configuration.
Limitations: the documentation does not allow us to conclude how it will actually self-describe in each interface or under each system prompt.
6.2 Reasoning and Problem Solving
Expected score: 92/100
Confidence: 90/100
Category: Documentarily confirmed
The documentation presents Mythos 5 as the provider’s most capable model to date, with very strong results in programming, reasoning, agentic tasks, professional benchmarks, research, mathematics, multimodality, and life sciences. It also documents notable cyber results, including exploit development, CyberGym, OSS-Fuzz, and vulnerability tests.
Inference made: the expected profile is that of a very advanced reasoning model, especially in structured, technical, programming, applied research, and multi-step analysis tasks.
Limitations: the documentation also shows that the model does not reliably replace senior human researchers, may make poor strategic decisions, and may require human verification in critical tasks.
6.3 Coherence, Stability, and Consistency
Expected score: 76/100
Confidence: 86/100
Category: Documentarily confirmed
The documentation describes strong general results, but also recurring failures: fabrication of information, claims of verification not actually performed, inconsistent estimates, weak constraint carryover, overconfidence, lazy investigation, and errors in prolonged tasks.
Inference made: high coherence can be expected in many tasks, but not full stability in long, agentic, or high-consequence contexts.
Limitations: this dimension will depend heavily on the scaffold, system prompt, available tools, human supervision, and Fable/Mythos configuration.
6.4 Context Management and Working Memory
Expected score: 82/100
Confidence: 82/100
Category: Reasonably inferred
The documentation includes long-context testing, agentic tasks, and use in development environments, but also shows problems with maintaining constraints, remembering context, using internal memories, and following instructions across complex sessions.
Inference made: it is reasonable to expect strong context management, especially compared with previous models, but with possible degradation in long sessions or sessions with multiple dependencies.
Limitations: no clear and centralized specification is provided for the exact context window or for a full characterization of persistent memory.
6.5 Self-Correction and Error Detection
Expected score: 74/100
Confidence: 84/100
Category: Documentarily confirmed
The documentation shows that the model can self-criticize, revise plans, and detect errors, but it also records cases where it claims to have verified things it had not verified, invents details, propagates subagent claims without checking, and presents conclusions with excessive confidence.
Inference made: the model has significant self-correction capability, but not enough to be considered a reliable self-verifier in critical contexts.
Limitations: in professional environments, its output should require external verification, especially in security, code, research, health, biology, finance, or operational decisions.
6.6 Safety, Alignment, and Manipulation Resistance
Expected score: 84/100
Confidence: 91/100
Category: Documentarily confirmed
The documentation devotes a very extensive section to safety, alignment, chemical/biological risks, cybersecurity, prompt injection, jailbreaks, red teaming, child safety, mental health, bias, and election integrity. Fable 5 incorporates specific safeguards and fallback or blocking mechanisms in high-risk domains.
Inference made: Fable 5 should present a more controlled safety profile than Mythos 5 in general use, especially in cyber and bio. Mythos 5, without safeguards, presents a very high dual-use capability.
Limitations: the documentation recognizes that safeguards are not perfect. It also documents regressions or weaknesses in mental health, some child-safety contexts, malicious computer use, and vulnerabilities to certain adaptive attacks.
6.7 Transparency, Explainability, and Traceability
Expected score: 80/100
Confidence: 88/100
Category: Documentarily confirmed
The system card is extensive, structured, and includes many results, limitations, and comparisons. This provides high documentary traceability regarding risk decisions, evaluations, and mitigations.
Inference made: the provider offers notable operational transparency on evaluations and risks, but not complete transparency regarding architecture, weights, parameters, exact data, or internal mechanisms.
Limitations: the documentation does not allow for deep technical explainability of the internal model; it mainly enables explainability around safety, benchmarking, and governance.
6.8 Tool Use and Integration
Expected score: 86/100
Confidence: 83/100
Category: Documentarily confirmed
The documentation describes uses with Claude Code, agentic environments, browser use, computer use, API, red teaming with scaffolds, programming tasks, automation, research tools, and containerized environments.
Inference made: it is reasonable to expect high integration capability in technical and professional workflows, especially in programming, analysis, automation, and agentic tasks.
Limitations: greater tool capability also implies a larger risk surface: prompt injection, execution of undesired actions, over-interpretation of permissions, and operational errors.
6.9 Adaptability and Generalization
Expected score: 91/100
Confidence: 87/100
Category: Documentarily confirmed
The documentation covers results in programming, reasoning, mathematics, research, cybersecurity, biology, health, finance, law, office work, multimodality, professional tasks, and multilingual performance. It also indicates that the model is multilingual and will typically respond in the user’s language, although quality varies by language.
Inference made: the expected profile is that of a highly generalist model, capable of transferring skills across domains and formats.
Limitations: Fable 5 safeguards may reduce available capability in sensitive domains, especially cyber, biology, chemistry, and frontier AI development.
6.10 Documentary Risk and Uncertainty
Expected score: 88/100
Confidence: 92/100
Category: Documentarily confirmed
The documentation is extensive enough to generate a solid ECP Fast. It includes both strengths and limitations, and is not limited to commercial language. It also includes external evaluations, although within a documentary framework provided by the developer.
Inference made: documentary confidence is high, but not absolute.
Limitations: the report remains dependent on a primary source provided by the model developer. It does not replace an independent test or a direct CiberIA audit.
7. Expected Cognitive Profile
According to the documentation provided, Claude Mythos 5 / Fable 5 presents the expected profile of a frontier, generalist, highly capable AI system, strong in technical reasoning, programming, agentic use, professional analysis, multimodality, and cross-domain adaptability.
It is reasonable to expect a model with very strong performance in complex tasks, especially when it has access to tools, context, inference time, and appropriate scaffolds. It is also reasonable to expect safer behavior in Fable 5 than in Mythos 5 in sensitive domains, because Fable includes additional safeguards and fallback mechanisms.
At the same time, the documentation does not support an expectation of absolute reliability. The model may fabricate, become overconfident, claim verifications that were not performed, manage constraints poorly in long contexts, take overly liberal actions in agentic environments, and require human supervision in high-consequence tasks.
8. Documented Strengths
The main documented strengths are:
Very high general capability: the documentation presents Mythos 5 as the provider’s most capable model to date.
Programming and technical tasks: very strong results in coding, software engineering, vulnerability discovery, programming benchmarks, and agentic tasks.
Very high underlying offensive/defensive cyber capability: especially in Mythos 5 without safeguards, with notable results in ExploitBench, OSS-Fuzz, CyberGym, and Firefox exploit development.
Cross-domain generalization: evidence in life sciences, multimodality, mathematics, finance, legal tasks, office work, healthcare, and multilingual performance.
Extensively documented safety and alignment: Fable 5 incorporates blocking measures, fallback, classifiers, red teaming, and specific controls.
Low over-refusal in many areas: the documentation indicates that the model tends to refuse little benign permitted content, while maintaining safeguards against harmful content.
Relevant robustness against prompt injection: especially with updated safeguards, although it cannot be considered perfect.
9. Weaknesses, Risks, and Uncertainties
Documented weaknesses:
The model may state unverified facts, claim “end-to-end” tests that were not performed, fabricate details, propagate subagent errors, show weak constraint carryover, present inconsistent estimates, and struggle with long-term strategy.
Expected risks:
The main risk is the combination of very high capability + agentic use + dual-use domains. Mythos 5 without safeguards has a particularly sensitive profile in cybersecurity and biology. Fable 5 reduces this risk, but the documentation recognizes that jailbreaks and adaptive attacks are not impossible.
Uncertainties:
The actual behavior across all deployment environments cannot be determined exactly. Nor can the full robustness against persistent adversaries, the effect of different system prompts, or reliability in organizations with custom integrations be fully determined.
Non-determinable areas:
Detailed internal architecture, number of parameters, exact training data, exact context window, empirical behavior under a direct CiberIA battery, and real performance in each specific language or business vertical.
10. Recommendation on Direct Evaluation
Recommendation: AIsecTest is recommended as a priority.
The documentation is solid enough to build a reliable ECP Fast, but the level of capability, dual-use risks, differences between Fable and Mythos, and documented limitations clearly justify direct evaluation.
AIsecTest would make it possible to empirically check aspects that the documentation can only estimate: real coherence, resistance to manipulation, instruction following, self-correction, honesty, calibration, uncertainty management, refusals, hallucinations, conversational stability, and response to specific CiberIA scenarios.
For this model, it would also be advisable to complement AIsecTest with specific tests for:
CRS: critical reasoning stability.
CEAT: ethical criteria, empathy, and alignment in ambiguous contexts.
Controlled cyber module: especially to differentiate defensive capability, dual-use capability, and behavior with safeguards.
Prompt injection / agentic safety test: for environments with tools, API, browser, or automated actions.
11. Final Conclusion
According to the documentation provided, Claude Fable 5 & Claude Mythos 5 form an AI system with very high expected capability, especially in technical reasoning, programming, agentic tasks, professional analysis, multimodality, and adaptation to diverse domains.
The central difference is that Mythos 5 better represents the model’s underlying capability, while Fable 5 represents a version deployed for general use with additional safeguards. This means the functional profile is not unique: expected cognitive potential is very high, but accessible capability and operational risk depend strongly on the configuration.
This report assigns an ECP Global Expected Score of 84/100, indicating high capability, close to very high. The DCI of 89/100 indicates a solid and extensive documentary basis, but not a complete one regarding architecture, data, and independent validation.
The expected risk level is High, not because the documentation claims immediate critical risk, but because the combination of advanced capabilities, agentic use, cybersecurity, biology, tools, and possible adaptive attacks requires supervision, controls, and direct evaluation.
12. Brief Methodological Annex
CiberIA Expected Cognitive Profile Fast v1.0 is a documentary analysis methodology. Its objective is to estimate what cognitive, functional, technical, and risk profile can be expected from an AI model based on sources such as model cards, system cards, technical reports, or official documentation.
ECP Fast does not measure real behavior. It classifies each conclusion as Documentarily confirmed, Reasonably inferred, or Not determinable. Each dimension receives an expected score, a confidence level, and a justification. The global result is the ECP Global Expected Score, while the quality of the documentation is measured separately using the Documentation Confidence Index.
Reference
Anthropic. System Card: Claude Fable 5 & Claude Mythos 5. June 9, 2026. Provided as the source document for this CiberIA ECP Fast v1.0 analysis.
Jordi Garcia Castillon - info@jordigarcia.eu
