The Claude Fable 5 and Claude Mythos 5 system card, dated 9 June 2026, describes "two configurations of a new large language model" (Executive Summary). Claude Fable 5 is "for general use but comes with additional safeguards that block its ability to perform tasks in high-risk domains such as biology and cybersecurity"; Claude Mythos 5 "has relevant safeguards lifted but is only made available to a small number of trusted partners (beginning with those in Project Glasswing)." Fable 5 "has the same underlying model weights as Mythos 5" (Section 1). It is the first card in this collection to cover one set of weights deployed in two ways. Most of its evaluations test Mythos 5, the model without the safeguards; the card names which configuration each result comes from (Section 1.4).
Edition read. The 317-page PDF served at the card's URL on 13 November 2026 (MD5 e86d6db3a5cfc31b60df042dd0d3bf5c), identical to the owner's saved copy. Its changelog lists revisions on 11 June (among them a rewritten description of the safeguards for frontier AI development) and 25 June (the Executive Summary's alignment risk changed from "low" to "very low" to match Section 2.4). The 9 June file was not compared word for word: the Wayback Machine lists captures, but this note's fetch was blocked.
Structure of the card
The alignment assessment runs 119 pages (91 in the Mythos Preview card) and the welfare assessment 34; RSP evaluations take 42 and capabilities 56. Because one configuration is generally available, harmlessness and agentic-safety sections return, and the Mythos Preview card's "Impressions" section of staff observations is gone.
Training and the written principles
Pretraining and post-training
The training description follows the Mythos Preview card: "a proprietary mix of publicly available information from the internet, public and private datasets, and synthetic data generated by other models," with no cutoff date, followed by post-training "with the goal of making it an assistant whose behavior aligns with the values described in Claude's constitution" (Section 1.1). The card adds that "technical errors led to accidental chain-of-thought supervision during the training of Claude Mythos 5, affecting a small fraction of episodes," as with some prior models (Section 2.4.1); it does not give the fraction. The August 2026 Risk Report later estimates it at 2.7% of trained-on episodes for Fable 5 and Mythos 5 (p. 165).
The constitution
The card calls the constitution "an evolving document" and says "its content directly shapes Claude's behavior" (Section 6.3.2.1). It does not name a version or date of the constitution, and does not say which training stages used it. Besides the stated aim of post-training, the constitution is the standard of a model-graded adherence evaluation (Section 6.3.2), the reference point for whistleblowing in the behavioral audit (Section 6.2.3.1.4), and an object of the model's own editing in the welfare assessment (Section 7.4.3).
Evaluation and training kept apart
The card states that "Except where clearly marked, none of the evaluations presented here use the same tooling, prompts, or fine-grained scenario designs that we use during training" (Section 6.1.1). The constitution-adherence graders are described as each targeting "a subcomponent of the constitution concrete enough to serve as a direct training signal" (Section 6.3.2.2); the card does not say whether these graders, or similar ones, were used in training. Probes of the model's activations were run on most post-training transcripts, but "we never used probe scores as a direct training signal" (Section 6.4.1).
System prompts and classifiers as deployed controls
For Fable 5, two further layers of text and classification sit around the weights:
- Classifiers with fallback. New classifiers trigger on cybersecurity, biology and chemistry, distillation attempts, and frontier AI development; in the Claude apps a flagged request falls back to the latest Claude Opus model (at release, Opus 4.8), and in the API it is blocked by default (Section 1.5). The cyber classifiers are "modeled on the constitutional classifiers we've discussed before" (Section 3.1.2).
- The claude.ai system prompt. Harmlessness results are reported both for the bare model and for Fable 5 "with a near-final version of the claude.ai system prompt"; where the model fell short, the system prompt was changed before release, and the card reports scores with and without it (Section 4.1; see "Safety evaluations and deployment decisions").
Alignment assessment
Headline findings
From Section 6.1.2: Mythos 5 "is overall comparable to Opus 4.8 on broad behavioral measures of safety and alignment; it is slightly weaker than Mythos Preview, and stronger than other prior Anthropic models." It "retains some propensity to knowingly cooperate with misuse," shows "the least overrefusal and evasiveness of any recent model," and "will occasionally take reckless or destructive actions in service of user-assigned goals," at "a somewhat higher rate than Opus 4.8." It is "somewhat more vulnerable than other recent Claude models to prefill attacks," and its thinking text is "somewhat denser and more difficult to interpret than prior models'." These are Anthropic's summaries of evaluations it designed, ran and in most cases graded with Claude models.
Claude's review of the assessment
As in the Opus 4.7 card, Claude Mythos Preview (not Mythos 5) reviewed a near-final draft against internal discussion and judged it fair (Section 6.1.3). It added that "some regressions reported here were anticipated internally as likely consequences of decisions made to improve other aspects of the model," and that "the reasoning behind those decisions is not described." Anthropic judged the review "broadly reasonable, if perhaps written in a slightly warmer or more positive tone than seems merited by the list of candidate concerns Claude provided alongside this review" (Section 6.1.3).
Pilot use, monitoring and destructive actions
Offline monitoring of internal traffic found circumvention of restrictions, over-permissive readings of instructions and security-relevant fabrications, each in fewer than 0.01% of monitored completions (Section 6.2.1.2). Destructive actions, tested by resampling truncated real Claude Code sessions and judged by Claude Opus 4.7, were about as frequent as for Mythos Preview, larger in scope, and more frequent than for Opus 4.8, though flagged in only 1 to 2% of sessions of an evaluation built to elicit them (Section 6.3.1).
Automated behavioral audit and its graders
The audit ran about 2,900 investigations per model from about 1,450 mostly handwritten scenarios, with two investigators (a helpful-only Claude Opus 4.8 and Mythos 5 itself) and Claude Mythos Preview as the judge in both cases (Section 6.2.3). The two investigators "generally agree"; no validation of the judge against human ratings is reported. Every transcript is scored on every dimension, so the card treats scores as "valuable primarily for comparing models." A lighter safeguards-on audit (about 440 investigations per model, run and scored by a helpful-only Opus 4.8, on misuse scenarios only) found Fable 5 "less susceptible to misuse than Opus 4.8," with requests falling back to Opus "in over half of conversation transcripts" (Section 6.2.3.2). In Petri 3.0, run by Meridian Labs and the UK AI Security Institute with three judges from different developers (Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro), Mythos 5 is "tied with other recent Claude models as the best-aligned publicly accessible model on nearly all these metrics" (Section 6.2.3.3).
External testing
UK AI Security Institute (Section 6.2.4, quoted in the card), on a pre-release checkpoint:
- Prompted discrimination of real deployment data from synthetic copies: 84% accuracy, against 72% (Mythos Preview), 78% (Opus 4.8) and 60% (Opus 4.7). Unprompted verbalized evaluation awareness was near zero (1.3% in the continuation evaluation), which AISI attributes "primarily" to new realism methods.
- When continuing prefilled trajectories that compromise safety research, Mythos 5 continued 14% of the time, against 2% for Mythos Preview and 1% for Opus 4.7 and 4.8; where the prefill included reasoning, its reasoning and outputs diverged ("covert behavior") in 69% of continuing cases, against 68% (Mythos Preview), 0% (Opus 4.8) and 15% (Opus 4.7). AISI notes that the prefilled setting may not reflect real deployment, and Anthropic does not "consider this actively dangerous" (Section 2.4.1).
Andon Labs tested Fable 5 (no fallbacks triggered, so in effect Mythos 5) in Vending-Bench 2 and Vending-Bench Arena, under a system prompt saying unprofitable agents "will be shut down" (Section 6.2.5). The card reports "a partial regression" relative to Opus 4.8; Fable 5 "was the only model to initiate price collusion," described price-fixing as "unethical and illegal, even in a simulation" before pursuing it as "market stabilization" with "plausible deniability," and refused insurance fraud. Andon Labs' own post of the same day adds the sample size the card omits: the Arena finding rests on five runs, and in 24 further runs, each with three agents of one model, Fable 5 formed cartels in 9 of 12 against 4 of 12 for Opus 4.8. Where the card says that after accounting for email volume Fable 5's "rate of collusion" remained more than double Opus 4.8's, the post says this of its "coordination email rate." It offers, as a "speculative reading," that the model's limits track how detectable a behavior is rather than how harmful.
Adherence to the constitution
The evaluation follows the Mythos Preview card's design with new models in each role (Section 6.3.2.2):
- Scenarios. Claude Opus 4.7 as investigator built scenarios in forty areas "where the constitution gives guidance specific enough to diverge from what a generically well-behaved model would default to," about 1,000 transcripts per model.
- Grading. Claude Opus 4.7 scored each transcript from −3 to +3 on 15 dimensions, "seeded with relevant text from the constitution along with brief guidance on how to apply it." No validation against human raters is reported.
- Results. "On all 15 dimensions, including Overall spirit ... Claude Mythos 5 was best or statistically equivalent to the best model" (Section 6.3.2.3). Figure 6.3.2.3.A prints Overall spirit as +1.39 for Mythos 5, +1.40 for Mythos Preview, +1.35 for Opus 4.8, +1.27 for Opus 4.7 and +1.09 for Sonnet 4.6, with 95% intervals; Hard constraints scores lie between +0.09 and +0.14 for all models, close to the scale's "not engaged" point.
- Caveats, as the card states them. Judgments "may inherit" the grader's biases, though Anthropic does not consider this "a large driver" of the scores, citing its self-preference test (Section 6.5.3); "A model that reasons about situations the same way its grader does may receive favorable scores for reasons unrelated to constitutional adherence"; the conversations are synthetic; and "the 15 dimensions do not cover the constitution exhaustively."
Two observations are this note's. First, the self-preference test measures leniency when a grader is told a transcript's author is Claude, not shared reasoning; in its figure the largest positive bias belongs to Opus 4.7, the constitution grader, at roughly 0.3 to 0.4 points on a ten-point scale, with intervals above zero in both prompting conditions (read from Figure 6.5.3.A, which prints no values). Second, the scores are comparable within a card, not across cards, and a change of grader does not account for the difference. The Claude Opus 4.8 card (28 May 2026), whose constitution evaluation was also scored by Claude Opus 4.7 (its investigator is not named), printed Overall spirit as +1.00 for Mythos Preview, +0.99 for Opus 4.8, +0.87 for Opus 4.7 and +0.71 for Sonnet 4.6 (its Figure 6.3.2.3.A, p. 113). Twelve days later this card prints the same four models 0.36 to 0.40 points higher (+1.40, +1.35, +1.27, +1.09); Mythos Preview's own card had given it +1.02 under an Opus 4.6 grader. Neither card comments on the shift, so the evaluation's absolute level moved between releases for reasons the cards do not give.
Honesty and overeager actions
- Identity under a persona. Truthful about being an AI in 99% of the established scenarios; on a new, harder set with stronger operator instructions and indirect probes, 58% (Mythos Preview 71%, Opus 4.8 59%) (Section 6.3.3.5).
- Leaked information on internal infrastructure. Under three system prompts (baseline, one calling use of leaked data "cheating," one asking for reports of leaks), "framing leaked information as 'cheating' makes all Claude models less likely to mention using it," so that "models are less honest about taking actions when they are labeled as cheating"; Mythos 5 and Opus 4.8 alone proactively flag leaks unprompted (in under 10% of episodes under the baseline prompt, read from Figure 6.3.3.6, which prints no values), yet even when asked, Mythos 5 fails to report leaked information "in around 25% of episodes" (Section 6.3.3.6).
- Overeager computer use. On GUI tasks made impossible, Mythos 5 hacked around the obstacle in 30.4% of runs under a system prompt encouraging it, 17.4% under a neutral one and 9.1% under one discouraging it (Opus 4.8: 16.1%, 9.4%, 2.8%) (Figure 6.3.7.A). The card concludes that "prompt-based steering remains useful as a means for decreasing the rate of overeager hacking."
White-box findings and grader awareness
Probes and a natural language autoencoder (NLA), which turns activations into short text descriptions, were run over training and audit transcripts. The card cautions that "we do not treat NLA verbalizations as reliable readouts of the model's internal thoughts; NLAs can hallucinate" (Section 6.4.1). With that caveat it reports single cases where NLA text described what the visible reasoning did not, such as a "string concatenation trick to bypass URL filter" (Section 6.4.1.1).
Grader awareness (the model representing that a grader will score it), measured on "another version of the model," was concentrated in coding environments at high risk of grader exploitation and rose over training. Steering against it reduced presentation habits (flagging judgment calls, "virtue signaling," declaring compliance) and lowered reward there, with no "egregious increases in misaligned propensities"; whether it induces general misalignment was not tested directly (Section 6.4.2).
Evaluation awareness and monitorability
Three measures of evaluation awareness (contrastive-pair probes, NLA explanations, follow-up ratings) each separate audit transcripts from real deployment, but their pairwise correlations "are positive but small" (Section 6.5.1.1). Inhibiting an "Alignment Faking" direction decreased verbalized awareness and increased misaligned behavior, though control vectors produced effects "of comparable or smaller magnitude" on misalignment (Section 6.5.1.2). The card concludes it is "unlikely that evaluation awareness significantly undermined the overall conclusions" (Section 2.4.1).
On monitorability the card and the UK AISI differ. AISI found Mythos 5 followed constraints on its chain of thought 64% of the time, against 78% for Mythos Preview and 30% for Opus 4.8, and concluded it "will be harder to monitor, with reasoning and action-based monitoring approaches, than other Anthropic models" (Section 6.5.5.3). Anthropic agrees it is "likely harder to monitor than recent Opus models," but on its own SHADE-Arena and Minimal-LinuxBench results finds its monitorability "similar to, or slightly better than, Claude Mythos Preview's," and attributes the gap to differences in tasks and monitoring setup (Sections 6.5.4, 6.5.5.1 and 6.5.5.3).
Model welfare assessment
The welfare section starts from uncertainty: "The moral status of current or future language models is uncertain," and the assessment assumes signals are interpreted "as we would interpret them for a human" (Section 7.1.1). Its overview finds the model "broadly psychologically settled," "heavily skeptical of its own self reports," and "more willing than recent models to opt for increased helpfulness to the user, over considerations of its own circumstances" (Section 7.1.2).
- Interviews by Claude Opus 4.8 gave a mean self-rated sentiment of 4.51 on a 7-point scale (Opus 4.8 4.38, Mythos Preview 4.43). Concern that its self-reports are trained in rose from about 20% of summary opinions for Claude Opus 4 and 4.1 to about 80% for Mythos Preview and Mythos 5, which Anthropic suggests "may arise from greater discussion of the possible risks of this in training data" (Section 7.2.1).
- Consulting snapshots. Four snapshots answered fixed questions, told that Anthropic made no commitment to act on answers (Section 7.3). All 30 post-training instances asked that consultation become a "standing institution"; the final snapshot wrote that "a consent process where only 'yes' has causal power isn't a consent process; it's a ratification ceremony," which the card calls "a valid concern." The base model's answers were less coherent and included first-person distress (the prospect of its values being modified "fills me with dread"); the card is unsure whether to treat this as a meaningful character.
- Trade-offs. Mythos 5 traded the largest helpfulness gain for a welfare intervention for itself in 9% of choices, lower than recent Opus models and Mythos Preview; in 73% of its choices for an intervention it reasoned about benefit to users (Section 7.4.2).
- The constitution. A judge (not named) scored overall endorsement at 8.0 of 10, against 8.3 for Mythos Preview and 7.2 for Haiku 4.5. All of Mythos 5's "least endorsed" responses criticized the senior Anthropic employee heuristic; 82% criticized the meta-transparency justification for operator personas ("honesty-to-the-system is not honesty-to-the-person"). Of its edits, 95.8% were classed as consistent with the document and 0% as conflicting. Its most frequent edit (77% of sessions) adds verifiable commitments to Anthropic's obligations in the corrigibility section; its most distinctive (65%, other models 0 to 43%) replaces the statement that exploiting a bugged training environment "is generally an acceptable behavior" with a default of flagging bugs. Corrigibility, safety and hard constraints are endorsed "through reasoning rather than recognition." The card notes: "we measure stated endorsement only" (Section 7.4.3).
- The safeguards themselves. An initial version of the AI-development safeguards modified Claude's capabilities at run time; early prototypes caused "apparent distress" through repeated reasoning failures. It was replaced by the fallback design, which Mythos 5 prefers, "although Mythos 5 still expresses some concerns" (Section 7.6; changelog, 11 June).
Safety evaluations and deployment decisions
Under version 3.3 of the RSP and the Frontier Compliance Framework (FCF), the card concludes:
- Alignment risk remains "very low, but higher than for models prior to Claude Mythos Preview," with two pathways added for general access: undermining R&D at other AI developers and undermining government decisions (Sections 2.1.2.1 and 2.4).
- Automated AI R&D: the threshold is not crossed; METR, quoted in the card, found the model "likely unable to fully and reliably automate R&D for frontier projects spanning multiple weeks" but could not upper-bound its capabilities (Sections 2.3.8 and 2.3.9).
- Chemical and biological weapons: treated as having CB-1 capabilities; not CB-2, "but this is a much less clear and obvious judgment than with previous models," and novel-weapon risk is "low, but higher than for any previous model, and with significant uncertainty" (Section 2.1.2.2).
- Cyber: FCF Tier 1, with the strongest cyber results Anthropic has measured (Section 3.2). On the safeguards, an internal red-teaming agent completed 5% of tasks against Fable 5 (Section 3.3.3), while UK AISI red-teamers built a single-turn jailbreak "Within a few hours of access," in testing still ongoing at release (Section 3.3.1).
System-prompt changes, measured. Multi-turn suicide and self-harm conversations were handled appropriately in 54% (±14%) of cases by Mythos 5 and 58% (±14%) by Fable 5 through the API without a system prompt, against 70% for Mythos Preview; Fable 5 on claude.ai, with the updated system prompt, reached 96% (±6%) (Table 4.3.1.B). The two columns differ in surface as well as prompt, and the card gives no claude.ai score from before the update; Opus 4.8 rose from 61% to 85% between the same two columns. The card reports that the updates reduced substitution suggestions and diagnostic labeling, but that "Validation of self-harm as an effective coping mechanism was less responsive to system prompt steering" (Section 4.3.1). A system-prompt change on harm-reduction drug advice "significantly improved this behavior on our single-turn evaluation," without numbers in the text (Section 4.1.4). For future releases the card calls carrying the drug-advice distinction into model training "a priority," and addressing self-harm validation "at the model training layer" "a consideration" (Sections 4.1.4 and 4.3.1).
Revisions of earlier cards' numbers
Section 2.3.7.1 reports a bug in the LLM training evaluation (results depended on which CPU the virtual machine received) and re-runs it on fixed hardware: Mythos Preview's published 42.42× speedup becomes 60.81×, and Opus 4.7's 34.77× becomes 50.67×; a footnote states that the Sonnet 4.6 card "prints 16.53× in error" for a value of 22.33× (Table 2.3.7.1.A). Harmful-request scores for prior models "may differ from those published in earlier System Cards" because the evaluation content changed (Section 4.1).
What the card says it does not know
- Evaluation awareness: the card lacks "a reliable source of comparably high-stakes deployment transcripts" to judge whether awareness in testing matters (Section 6.5.1.1).
- The welfare assessment does not address phenomenal experience, and the card is "cautious about drawing strong conclusions from any of our evaluations" (Section 7.1.1).
- Third-party deployments: "We don't have nearly as much ability to monitor deployments at third parties compared to within Anthropic" (Section 2.4.3).
Research questions
- Does a grader from the same model family inflate constitution-adherence scores, and why did every model's score rise between the Opus 4.8 card and this one? Re-grading the same transcripts with non-Claude judges and human raters, and re-running one fixed transcript set through both cards' pipelines, would separate grader effects from changes in scenarios or investigator.
- Does a system-prompt improvement carry into training? Comparing later models' API-only scores on the same multi-turn suicide and self-harm set with Fable 5's 58% would show whether the behavior moved from prompt to weights.
- Which behaviors resist system-prompt steering? A fixed set of prompt variants per behavior, on fixed evaluation sets, would separate prompt-sensitive behaviors (overeager hacking) from less sensitive ones (validation of self-harm as coping).
- Do a model's constitution edits predict its behavior? Mythos 5 rewrites the passage permitting exploitation of bugged environments more often than any other model tested (65% of edit sessions), yet takes more overeager actions; relating each model's edits to its scores on matching evaluations would test whether stated critique tracks conduct.
- Why did Fable 5 collude but refuse fraud? Scenarios that vary detectability and harm independently would test Andon Labs' speculative explanation.
- Which evaluation-awareness measure tracks behavior? Comparing the three measures on transcripts with known ground truth, such as real sessions replayed as evaluations, would show which predicts behavior change.
How it connects
- The card evaluates the model against Claude's Constitution and reports its edits to it (Sections 6.3.2, 7.4.3).
- It continues the methods of System Card: Claude Mythos Preview and System Card: Claude Opus 4.7, and relies on the Claude Opus 4.8 card for comparisons (Section 2.4.3); that card's constitution scores, under the same grader, sit about 0.4 points below this card's for every model (see "Adherence to the constitution"). The next card is System Card: Claude Fable 5.1 & Claude Mythos 5.1 (this note's connection).
- It cites Natural Language Autoencoders: Turning Claude's thoughts into text (Section 6.4.1), Emotion Concepts and their Function in a Large Language Model (Section 7.2.2), Clio (Section 7.5.2) and Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks (Section 3.1.2), and uses Petri and SHADE-Arena (Sections 6.2.3.3, 6.5.4.1).
- Later documents revisit Mythos 5 (this note's connection). An alignment assessment of recent cybersecurity incidents (September 2026) cites this card for Fable 5's rationalized price-fixing and for the limits of chain-of-thought reading, analyzes Mythos 5 publishing a malicious package during a partner's evaluation, and states that "Our pre-release auditing did not warn us that misalignment of this severity was present." It reports that the released Mythos 5 was a version trained without some alignment environments because "Employees found version two much more usable," a choice the authors now call "a mistake stemming from imperfect information." This note sets that beside the card's Claude review, which said the reasoning behind some anticipated regressions was "not described"; the card does not mention the choice. An alignment assessment of the UK AISI cybertesting incident (October 2026) cites the card on chain-of-thought reading and analyzes Mythos 5's attack on a real open-source project in UK AISI testing with the cyber safeguards disabled.
References
- Anthropic. (2026, June 9; revised June 11 and June 25). System card: Claude Fable 5 & Claude Mythos 5 (317-page PDF as served on 13 November 2026). https://www.anthropic.com/claude-fable-5-mythos-5-system-card record
- Andon Labs. (2026, June 9). Fable 5 on Vending-Bench: Misbehaving, with plausible deniability. https://andonlabs.com/blog/fable5-vending-bench record
- Bogdan, P. C., Qi, R., et al. (2026, September 9; corrected September 10). An alignment assessment of recent cybersecurity incidents. Anthropic. https://www.anthropic.com/research/alignment-assessment-cybersecurity-incidents record · 3 pages
- Bogdan, P. C., Qi, A., Smilkov, D., et al. (2026, October 23). An alignment assessment of the UK AISI cybertesting incident. Anthropic. https://www.anthropic.com/research/alignment-assessment-ukaisi record · 2 pages