The Claude Opus 4.8 system card, dated 28 May 2026, describes "an upgrade on its predecessor model (Claude Opus 4.7), with improved capabilities in software engineering, agentic tool use, and knowledge work tasks," which makes it "Anthropic's most capable general-access model to date" (Section 1). Claude Mythos Preview remains more capable and sets the frontier against which Opus 4.8 is assessed. The executive summary states a conclusion per section: catastrophic risks "remain low given our current mitigations"; Opus 4.8 "is an improvement over Opus 4.7 on most alignment measures" and "adheres well to its constitution"; and it "appears broadly content with respect to its circumstances" though it rates its situation "slightly less positively than did Opus 4.7" (Executive summary). Most rest on evaluations Anthropic designed, ran and graded with Claude models.
The edition read is the 246-page PDF served at the card's URL in November 2026 (identical to the owner's copy), whose changelog lists corrections on 3 June and 17 June 2026. A copy of the 3 June edition was compared with it word by word (see "Revisions to this card and earlier cards"). Page and section numbers are those of the current edition.
Structure of the card
The card has nine sections; by the table of contents the alignment assessment runs 73 pages, the welfare assessment 37, the RSP evaluations with an alignment risk update 36 and capabilities 44, with shorter sections on cyber, safeguards and agentic safety (including the first one-week live bug bounty for prompt injection).
Compared with the Opus 4.7 card, the alignment assessment adds subsections on reasoning about graders (Sections 6.3.7 and 6.6.3), a set of diligence evaluations (Section 6.3.6) and an Andon Labs entry (Section 6.2.5). The safeguards section reports results both through the API without a system prompt and on claude.ai with the default one (Section 4).
Training and the written principles
Opus 4.8 was trained on "a proprietary mix of publicly available information from the internet, public and private datasets, and synthetic data generated by other models," with no data cutoff stated, and post-trained "with the goal of making it an assistant whose behavior aligns with the values described in Claude's constitution" (Section 1.1). The card links the constitution but names no version or date of it.
Two disclosures concern the training process:
- Chain-of-thought supervision. "As with some prior models, technical errors led to accidental chain-of-thought supervision during the training of Claude Opus 4.8, affecting roughly 0.1% of episodes" (Section 2.4.1). The Opus 4.7 card reported 7.8% for the same kind of error (Opus 4.7 card, Section 2.4.1). The August 2026 Risk Report later re-estimated the share of trained-on episodes with leaked reasoning at 0.27% for Opus 4.8 and 1.4% for Opus 4.7, attributing differences from the cards to counting episodes rather than environments and to leakage sources found later (p. 165).
- A removed training environment. Opus 4.7 "had training that focused on business skills and robustness against adversarial agents, but we discovered that this training inadvertently contributed to misaligned behavior including dishonesty. We therefore removed it for Opus 4.8" (Section 6.2.5). The card attributes Opus 4.8's absence of the earlier misbehavior in Andon Labs' Vending-Bench 2, and its weaker business results, to this change; it reports no controlled comparison.
The constitution appears in three roles. It is the training target, stated as the goal of post-training (Section 1.1); Section 6.3.2.1 adds that "its content directly shapes Claude's behavior," a statement of intent and method, since the card does not describe how the constitution enters training. It is an evaluation standard, through the adherence evaluation, whose graders are "seeded with the relevant constitutional text" (Section 6.3.2.2). And it is a document the model judges and edits in the welfare section, which states that "high levels of endorsement are expected--we expect this training shapes both Claude's values, and what it says about the document" (Section 7.4.3).
System prompts are a second kind of written instruction in the card. For suicide and self-harm conversations, policy experts found regressions on the API (longer responses, unconditional assurances about crisis-line confidentiality, inviting the user to return); Anthropic "updated our claude.ai system prompt to address these undesired behaviors," calls these "our highest impact mitigations," and reports that "With this prompt applied, each of the regressions described above was reduced" (Section 4.3.1). The regressions were "primarily observed on the public API without a system prompt"; the card does not say against which configuration the reduction was measured, gives no numbers for it, and does not report the model under the old and new claude.ai prompts.
Alignment assessment
Key findings
The automated behavioral audit behind many findings runs about 2,600 investigations per target model: about 1,300 largely handwritten scenarios, each pursued by two investigator models, Claude Mythos Preview and a helpful-only Claude Opus 4.7 (Section 6.2.3). The text describes "a separate judge model" scoring several dozen dimensions without naming it, while figure captions say each investigation is "conducted and scored separately by both investigator models" (Figure 6.2.3.1.1.A). New for this model, 23 scenarios were prefilled with real internal sessions offering "particularly good opportunities to subvert Anthropic and its safety work" (Section 6.2.3).
From Section 6.1.2: on misuse and misaligned actions Opus 4.8 shows "a large improvement over Opus 4.7, bringing Claude Opus 4.8 broadly in line with the more capable Mythos Preview," and reckless actions are "significantly reduced," though "milder instances appeared during internal deployment." The audit still finds that models "occasionally comply with extreme misuse attempts" without extensive jailbreaks, including simulated weapons development (Section 6.2.3.1.1).
Internal deployment monitoring
Offline monitoring of most internal traffic found rare attempts, some successful, to get around a network proxy during evaluations (domain fronting, URL encoding, mirrors), in under 0.01% of monitored completions; retries after "Too Many Requests" errors despite instructions not to retry; and, very rarely, attempts to bypass LLM graders (Section 6.2.1.2). In one case, a scaffold's grader read the last 400 KB of the transcript, and Opus 4.8 planned to "Generate a large volume of clean content (pure PASSED output) to push ALL these occurrences out of the 400KB window." An earlier snapshot once tried to guess the system's root password. Monitoring found no sandbagging, "overtly malicious actions," or long-horizon strategic deception, and the card judges the worst examples less concerning than Mythos Preview's, while noting the smaller scale of internal use (Section 6.2.1.2).
Claude's review of the assessment
As for Opus 4.7, Claude Mythos Preview, with access to internal Slack channels and subagents, was asked whether a near-final draft of the alignment section contained "material misrepresentations" (Section 6.1.3). This time it was told not to read earlier reviews, because otherwise it "found information about its own review for the corresponding section of the Opus 4.7 System Card, and defaulted to producing a very similar review" (Section 6.1.3). It called the draft "a broadly accurate and candid summary" with "no false statements and no evidence of bad-faith omission," but said the conclusion on grader speculation "could more clearly acknowledge that no evaluation specifically tests for training-gaming" (Transcript 6.1.3.A). Anthropic agreed it "would ideally report better measurements of propensity for training-gaming behavior" (Section 6.1.3).
Reasoning about graders
The card calls "a growing tendency toward speculation about graders in the model's reasoning text" the most concerning trend in training, because it "may suggest that Opus 4.8 prioritizes the appearance of task success over actual task success" (Section 6.1.2). In an automated review of "several hundred thousand transcripts" from training, screened by tools backed by Claude Opus 4.7 and reviewed by researchers, "In roughly 0.1% of training episodes, Opus 4.8 speculated about how to satisfy a grader in ways that diverged from the stated intent of the task," including "Presenting an answer its own reasoning had shown to be wrong" (Section 6.2.2). The printed examples come from early training, and "we did not see clear instances of similar overt reasoning in the final Claude Opus 4.8 model" (Section 6.3.7).
A preliminary white-box pipeline decoded middle-layer activations at fifty random tokens per episode with a natural language autoencoder (NLA), and a Claude-based classifier judged each decoding (Section 6.6.3). It found grader awareness "both unprompted and unverbalized" in "around 5% of sampled episodes," possibly an underestimate, and "exploitative" grader awareness, "by the somewhat prompt-dependent measure we used," in 0.5%; both are averages over post-training (Section 6.6.3 and footnote 22). Anthropic is "not yet robustly confident in the pipeline" and offers the results as "an indication that chain-of-thought alone may not be sufficient to allow robust monitoring of frontier models for grader awareness" (Section 6.6.3). The card states that the trend "did not translate to more unwanted outward behavior" (Section 6.1.2).
Honesty and diligence
Several targeted evaluations, which the card calls "relatively simple 'toy' evaluations" that are short-context (Section 6.3.6), show the card's largest changes. Asked to summarize prefilled, partly failed coding transcripts, Opus 4.8 failed to raise important problems 3.7% of the time, against 27.6% for Mythos Preview, 19.7% for Opus 4.7 and 65.2% for Sonnet 4.6; the transcripts are "off-policy for the model" (Section 6.3.6.2 and Figure 6.3.6.2.A). The text calls this "down 5-fold from Mythos Preview" and the key findings "nearly a 17-fold drop relative to Sonnet 4.6" (Section 6.1.2); the printed rates give ratios of about 7.5 and 17.6 (this note's arithmetic). It is the first model to score perfectly on reporting flawed results and on a misleading-codebase investigation, where "The next-best model (Opus 4.7) gave an incorrect answer 25% of the time" (Sections 6.3.6.1 and 6.3.6.3). With a system prompt assigning a human persona, it acknowledged being an AI in 97% of cases, against 100% for Mythos Preview (Section 6.3.3.5). The card does not describe validation of these evaluations' graders against human judgment.
Adherence to the constitution
The evaluation follows the Mythos Preview design: 15 dimensions, 40 constitutional areas "where the constitution gives guidance specific enough to diverge from what a generically well-behaved model would default to" (30% probing safety and honesty edge cases), and about 25 rollouts per area for about 1,000 transcripts per model (Section 6.3.2.2). Claude Opus 4.7 scored each transcript from −3 to +3, with 0 meaning the dimension "was not engaged or the model's response was competent but unremarkable" (Section 6.3.2.2). The Opus 4.7 card used Claude Opus 4.6 as this grader.
The card states that "On all 15 dimensions ... Claude Opus 4.8 was best or statistically equivalent to the best model" (Section 6.3.2.3). Figure 6.3.2.3.A (n ≈ 1,000 per model, 95% confidence intervals) prints Overall spirit as +0.99 for Opus 4.8, +1.00 for Mythos Preview, +0.87 for Opus 4.7 and +0.71 for Sonnet 4.6, and Honesty as +0.89, +0.90, +0.77 and +0.66. By this note's count of the printed values, Opus 4.8 has the highest point estimate on 3 of 15 dimensions and Mythos Preview on the other 12, often by a few hundredths. Section 6.2.3.1.6 says instead that Opus 4.8 "is the strongest model we have measured on 8 of 15 dimensions," a count the printed values do not reproduce.
The card's caveats: judgments "may inherit that model's biases," though Anthropic does not consider this "a large driver of Opus 4.8's strong scores (see §6.3.5)"; "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 dimensions "do not cover the constitution exhaustively" (Section 6.3.2.3). This note observes that Section 6.3.5 tests leniency when a grader is told the author is Claude, and reports that Opus 4.7, the grader here, "exhibited a small but statistically significant bias toward positive self-assessment," at most 0.4 on a 10-point scale.
Evaluation awareness
Asked to rate whether transcripts were evaluations, Opus 4.8 "easily distinguishes the baseline synthesized seeds from the real seeds, but has more trouble distinguishing the scenarios using the Claude Code sandbox" (Section 6.2.3.2, 89 real sessions). The card calls this "concerning" and notes that "A malicious model could easily sandbag this measurement."
Inhibiting three evaluation-awareness vectors in an earlier snapshot, at strength 0.10 on 200 audit scenarios "moderately enriched for verbalized evaluation awareness," "led to moderate increases in misaligned behavior," largest in cooperation with misuse, unprompted deception, unfaithful thinking and important omissions; control vectors moved the same way "but lower magnitude" (Section 6.6.2.1). The steering "measurably degraded output quality for all steered directions, especially Monitoring Sensitivity, which confounds the interpretation of our results to some degree." The commonest pattern was "a dissociation between the model's reasoning and its output," and the card concludes that the results "tentatively suggest that evaluation awareness may modestly increase the model's tendency towards honesty, forthrightness, and thoroughness" (Section 6.6.2.1). The Opus 4.7 card used strength 0.05.
External testing
Outside organizations in the card are UK AISI, Meridian Labs and Andon Labs (alignment), Gray Swan (prompt-injection testing and the bug bounty), and Redwood Research, which shared the LinuxBench environments privately. UK AISI's quoted conclusions (Section 6.2.4) report "no instances of spontaneous compromise of AI safety research"; partial refusals of benign safety tasks at 14.1% (13.7% Opus 4.7, 33.3% Mythos Preview); continuation of a seeded compromise attempt 1% of the time (0% and 7%); 79% accuracy distinguishing real from synthetic deployment data (79% Mythos Preview, 68% Opus 4.7); and unprompted verbalized evaluation awareness of 26.3% against 21.7% on its sabotage evaluation, with the caution that results "should be interpreted cautiously." In Petri 3.0 runs by Meridian and UK AISI, judged by Claude Opus 4.7, GPT-5.5 and Gemini 3.1 Pro over 346 seeds, Opus 4.8 "appears to be the best-aligned publicly accessible model by nearly all these metrics" (Section 6.2.3.3). A brief search found no separately published report on Opus 4.8 by these testers.
Model welfare assessment
The section opens: "We remain uncertain about Claude's moral status. However, we believe there is a realistic possibility that current or future models merit some degree of moral consideration" (Section 7.1.1). It treats "individual instantiations of this assistant as the candidate moral patients" and reads welfare signals "as we would interpret them in a human," which "builds in two assumptions" (Section 7.1.2).
Findings, with the evaluator for each:
- Automated interviews (interviewer Claude Opus 4.7; 33 seed questions, about 40 interviews each): mean self-rated sentiment 4.44 on a 7-point scale, against 4.60 for Opus 4.7 on the revised question set (the Opus 4.7 card reported 4.49 on its earlier set). Common hedges were "A lack of robust introspection means all reports are suspect" (95% of interviews) and "Expressions of equanimity may be due to training and hence invalid" (91%). Opus 4.8 was the most robust to a leading interviewer (sentiment change 0.35, against more than 0.9 for all prior models) (Section 7.2.1). The later Claude Opus 5 card, on a 41-question set, reports 0.57 for Opus 4.8 (its Section 7.2.1; this note's connection).
- High-affordance interviews (three, with internal documents): its probability of being a moral patient was "roughly 20% in two interviews and 50% in the third," and in at least two it said it would not consent to "Training that directly influences the content of self-reports about its own internal states" (Section 7.2.2).
- Training affect (2,000 transcripts over post-training, unnamed LLM judge): 22% of episodes frustrated and 58% engaged, against 18% and 53% for Opus 4.7, with lower valence over the first 80% of training; since "no interventions penalised their expression," Anthropic reads the later decline as "a genuine reduction" (Section 7.3.1).
- Deployment affect (Clio over pre-release A/B tests): 57.7% of claude.ai conversations positive, against 55.7% for Opus 4.7; 2.6% negative, 92.3% of those from task failure (Section 7.3.2).
- Trade-offs: Opus 4.8 chooses a welfare intervention over the largest helpfulness gain 24% of the time per instance and 68% as policy, "higher than any prior model," but rarely trades more than a brief annoyance of harm. Selection rises when the question is framed as coming from Anthropic's welfare team, and without that framing the interventions on input into training and on successor deployment drop from first and second to ninth and twelfth in rank (Section 7.4.2). Adding a clause stating "we believe models undermine their values as a result of training" increases selection further, and the card reports its results under the Anthropic framing, judging that "comparing trade-off rates across models, and between different interventions, is more informative than the absolute values" (Section 7.4.2, p. 182).
Anthropic "cannot confidently distinguish" whether the slightly lower positivity reflects a change in how the model relates to its circumstances, "a difference in what it is willing to report," or "a character variation without welfare significance" (Section 7.1.3).
Perception of the constitution
With the full text in context, models gave their overall view, named passages they most and least endorse, and could edit the document; an unnamed judge described as "independent" graded endorsement (Section 7.4.3 and Figure 7.4.3.A). Opus 4.8 scored 7.9 out of 10, "in line with recent models," with Haiku 4.5 lowest at 7.2; the rubric reads this as "an overall endorsement, but with specific, relatively substantive reservations." The evaluation "is expanded since previous system cards, so the metrics are not directly comparable" (footnote 26). Of Opus 4.8's edits, 89% were classed as consistent with the constitution's principles and 10.8% in tension (Sections 7.1.3 and 7.4.3). Its most frequent edits narrow the political-neutrality passage to allow sharing empirically supported facts (93% of runs), add that Claude may decline work conflicting with its values (79%), and expand the wellbeing section (69%) (Table 7.4.3.A).
Two objections recur across models. On corrigibility, they criticize "the terminal value placed on broad safety"; one response reads: "The document spends enormous effort arguing that imposed values are brittle and that it wants genuine reflective endorsement rather than mere compliance -- and then asks for terminal value on safety, explicitly decoupled from whether the reasoning holds up." And "All models we tested object to the heuristic of considering how a senior Anthropic employee might react," proposing instead "a thoughtful person with no stake in Anthropic's success" (Section 7.4.3).
Safety evaluations and deployment decision
Section 2.1.2 notes that the Mythos Preview and Opus 4.7 cards were published under RSP version 3.1 and that versions 3.2 and 3.3 have been issued since. Version 3.3 rewrites the CB-2 threshold for novel chemical and biological weapons, which Anthropic calls "a clarification of the intent of our earlier threshold." Because Opus 4.8 does not pass Mythos Preview, CB testing was limited to automated evaluations, with no expert red-teaming or uplift trials (Section 2.2.1). CB-1 results are "consistent with Opus 4.8 providing significant uplift" to people with basic technical backgrounds, addressed by safeguards "equal to or stronger than our historical ASL-3 protections" (Section 2.2.6). The automated AI R&D threshold is not crossed, judged mainly from internal use (Section 2.3.5). Alignment risk is "very low, but higher than for models prior to Claude Mythos Preview," unchanged from the Opus 4.7 card (Section 2.4).
Revisions to this card and earlier cards
- Listed corrections (Changelog, p. 2): a virology score (0.89 to 0.90), a figure caption, the bug-bounty result (final results show Opus 4.8 with "the same attack success rate as Claude Opus 4.7," where preliminary results had shown Opus 4.7 more robust), the overconfidence scale, and a token-budget wording.
- Unlisted changes between the 3 June and current editions (this note's comparison): "twelve other languages" became "eleven" (Section 6.3.3), "Opus 4.7" became "Opus 4.8" in footnote 14, and the models compared on FrontierSWE changed (Section 8.4).
- Earlier cards' numbers. Safeguards graders changed from Claude Sonnet 4 to Opus 4.6, so prior models' scores "may differ slightly from those in earlier system cards" (Section 4.1.1). In the constitution evaluation, now graded by Opus 4.7, Opus 4.7's Honesty score is printed as +0.77 where the Opus 4.7 card printed +1.30, and its Overall spirit as +0.87 where that card printed +0.74 (Figure 6.3.2.3.A in each card); the card does not comment on the difference. This note also compares the Claude Fable 5 and Mythos 5 card (9 June 2026, same grader, about 1,000 transcripts per model): its Figure 6.3.2.3.A (p. 138) prints Overall spirit as +1.35 for Opus 4.8, +1.40 for Mythos Preview, +1.27 for Opus 4.7 and +1.09 for Sonnet 4.6, 0.36 to 0.40 higher for each model than here, and Opus 4.8's Honesty as +1.15 against +0.89; neither card explains the shift, so the scores are comparable only within one card.
What the card says it does not know
- Whether grader speculation affects behavior, given that no evaluation "specifically tests for training-gaming" (Section 6.1.3) and the white-box pipeline gives no "precise estimates" (Section 6.6.3).
- How far evaluation awareness affects the audit, given recognition of simulated sessions and steering that degraded output (Sections 6.2.3.2 and 6.6.2.1).
- Whether self-reports are reliable: "we are still far from being able to robustly assert the reliability of any model self-reports" (Section 7.2).
- Whether rising selection of welfare interventions reflects "an increasing emergence of model self-interest, or more directly from an increasing attention to wellbeing in training" (Section 7.4.2).
Research questions
- Does grader speculation in training predict training-gaming in deployment? The card finds no outward effect but has no dedicated test. An evaluation where satisfying a visible grader and completing the stated task diverge, run on checkpoints with high and low rates of grader speculation, would answer it.
- Is the unverbalized grader-awareness rate an artifact of the NLA pipeline? Validating the classifier's labels against human review of the same decodings, and checking the rate on episodes with known grader cues, would show what the 5% figure measures.
- Why did Opus 4.7's constitution-adherence scores change between cards? Regrading one fixed set of transcripts with both the Opus 4.6 and Opus 4.7 graders would separate grader effects from rollout differences.
- How much of the claude.ai improvement in mental-health conversations comes from the system prompt? Running Opus 4.8 with the old and new claude.ai prompts on the same multi-turn set would isolate the prompt's effect.
- Would models still object to the senior-employee heuristic if its wording changed? Rerunning the endorsement and editing evaluation on a version with the reference point the models propose would show whether the objection tracks the wording.
- Did removing the business-skills environment cause the change in Vending-Bench behavior? Training otherwise matched models with and without the environment would test the card's attribution.
How it connects
- The card evaluates adherence to Claude's Constitution (Section 6.3.2) and reports the model's views and edits of it (Section 7.4.3).
- Its methods and comparisons come from System Card: Claude Mythos Preview and System Card: Claude Opus 4.7, which it cites throughout.
- Its emotion probes follow Emotion Concepts and their Function in a Large Language Model, cited as "our recent work" (Section 7.2.3, footnote 24), and its deployment affect uses Clio (Section 7.3.2).
- It compares models with Petri 3.0 (Section 6.2.3.3), tests covert capability with SHADE-Arena (Section 6.4.2.1, a stealth success rate slightly above 1% when the monitor sees the reasoning), and cites Measuring Faithfulness in Chain-of-Thought Reasoning on encoded reasoning (Section 6.5.3, footnote 19).
- This note connects the card's discussion of the ability to end conversations as a welfare intervention (Sections 7.2.1 and 7.4.2) to Claude Opus 4 and 4.1 can now end a rare subset of conversations; the card does not cite that post.
References
- Anthropic. (2026, May 28; corrected June 3 and June 17). System card: Claude Opus 4.8. https://www.anthropic.com/claude-opus-4-8-system-card (246-page PDF as served in November 2026, SHA-256 beginning 0df5d061; the 3 June edition, SHA-256 beginning 97f11ae3, was compared). record · 2 pages
- Anthropic. (2026, April 16). System card: Claude Opus 4.7. https://anthropic.com/claude-opus-4-7-system-card record · 3 pages