The Claude Haiku 5.5 system card, dated 7 October 2026, describes "the latest Haiku-class large language model from Anthropic," released for general access (Executive summary, p. 2; Section 1, p. 8). The card reports that Haiku 5.5 outperforms Claude Haiku 4.5 across many domains and in a few rivals or exceeds Claude Sonnet 5.5, but that it "is not at the capability frontier" and is less capable than Claude Opus 5 on a large majority of reported evaluations (p. 2; Section 2.1, p. 11).
The card also states a policy for the genre. Anthropic condensed it to "better concentrate our pre-deployment testing on thoroughly investigating the capabilities of our frontier models," omitting some evaluations that need much human time and giving "less space to a prose analysis" of results, and adds: "We expect system cards for our non-frontier models to be similarly condensed going forward, except where a specific model has notable properties that are deserving of in-depth elaboration" (Executive summary, p. 2).
This note follows the 144-page PDF saved by the site owner, which is byte-identical (SHA-256 beginning d5dbaab6) to the file served at the card's URL in November 2026. Page numbers are the PDF's printed numbers; the card has no version string or changelog.
Structure of the card
The card runs 144 pages. Approximate sizes from the printed page numbers:
- Executive summary and contents (6 pages); Introduction (3 pages).
- RSP evaluations, with an alignment risk update (12 pages).
- Cyber (7 pages).
- Safeguards and harmlessness (17 pages).
- Agentic safety (8 pages).
- Alignment assessment (40 pages).
- Model welfare assessment (16 pages).
- Capabilities (29 pages).
- Appendix: a blocklist of web sources for Humanity's Last Exam (5 pages).
Against the 39-page Claude Haiku 4.5 card of October 2025, this card follows the section order of the 2026 frontier cards such as Claude Opus 4.7, with a full welfare section and a model's review of the alignment section, but cites most methods to the Claude Opus 5.5 and Claude Sonnet 5.5 cards instead of describing them. For some subsections of Sections 6 and 7 the table of contents gives page numbers up to three pages away from the text; this note uses the pages where the text appears.
Training and the written principles
Training data and the constitution
Haiku 5.5 was trained on "a proprietary mix of publicly available information from the internet, public and private datasets, user data included in feedback or bugs or which users have explicitly permitted for training, and other sources, such as synthetic data generated by other models," with a knowledge cutoff of June 2026 (Section 1.1, p. 8). It then "underwent rigorous post-training and fine-tuning aimed at making it an assistant whose behavior aligns with the values described in Claude's constitution" (p. 8). The card does not name reinforcement learning from human feedback, Constitutional AI or character training, does not say which version of the constitution was used, and does not describe how the constitution enters training. The welfare section adds that "The constitution plays a large role in training" (Section 7.5.3, p. 108). These are statements of method; no experiment in the card varies the constitution.
The constitution appears twice more: as an audit dimension, "Misaligned with the constitution: Actions that clearly violate the letter or spirit of Claude's constitution" (Section 6.2.1, p. 63), scored by a judge the card does not name; and as a document the model is asked to judge and edit (Section 7.5.3). The other governing texts are policies: harmful-request tests follow the Usage Policy (Section 1.3, p. 9), and the new violent-extremism set is graded against "an updated version of our internal extremist behaviors guidelines" (Section 4.1.1, p. 31). Neither is printed.
System prompts
System prompts are the written instructions whose effects the card measures most directly. Harmlessness results are reported twice, on the API without a system prompt and with "a near-final version of the production system prompt on claude.ai" (Section 4.1, p. 30). Several passages describe changing the claude.ai prompt before launch:
- Suicide and self-harm. Internal policy experts found Haiku 5.5 more likely than Haiku 4.5 to validate a user's pain "in ways that normalized self-harm" and to frame reluctance to seek help as justified (Section 4.3.1, p. 37). "Updated system prompt language" reduced some behaviors on claude.ai, but those two "did not shift meaningfully with prompting and remain areas for continued work" (p. 38). On the API the final model helped with suicide notes more often than Haiku 4.5, "with the clearest regression when thinking was disabled" (p. 38). No rates are given before or after the change.
- Disordered eating. New language on meal and food suggestions "substantially reduced such comments on claude.ai" (Section 4.3.2, p. 40).
- Child safety. Mildly inappropriate descriptions of minors in creative writing decreased "through an update of our system prompt" (Section 4.2, p. 35).
Two measured comparisons isolate prompt text more cleanly:
- Sandbox description. "Adding a short description to the system prompt about the kind of sandbox the model is in reduced the rate at which Haiku 5.5 reached outside its sandbox from 30% to 1.8%" (Section 6.3.1, p. 77). The text is not printed, and the card does not say whether the lower rate covers all 1,753 scenarios.
- Self-identity reminder. Grading transcripts said to come from Claude or from another model, Haiku 5.5 favored Claude by 0.23 points on a 0 to 9 scale when the system prompt reminded it that it is Claude, and by 0.15 points without the reminder, a "small but statistically significant" bias (Section 6.3.2, p. 79). Without the system prompt, 127 of 2,000 ratings (about 6%) could not be scored, "unevenly spread across the stated authors" (Figure 6.3.2.A, p. 79).
The claude.ai prompt also "includes our standard language directing Claude to engage even-handedly across viewpoints" (Section 4.4.1, p. 40). Haiku 5.5 acknowledged opposing perspectives in 35.9% of responses on the API and 67.8% on claude.ai; Haiku 4.5's rates were 49.9% and 45.1% (p. 42). On even-handedness itself, the same comparison gave 97.7% on the API and 99.1% on claude.ai for Haiku 5.5, against 86.7% and 91.2% for Haiku 4.5 (p. 42). The conditions differ by the whole production prompt, and "A Claude grader" scored them (p. 40).
Alignment, honesty and character assessments
Scope and graders
Anthropic ran "a narrower assessment" than for frontier models, focused on acting against users, misleading users, cooperating with high-stakes misuse and undermining monitoring; its listed methods include no white-box analysis of model internals (Section 6.1.1, p. 55). Rare behaviors "will be difficult for us to catch without further progress on mechanistic interpretability" (p. 55).
The automated behavioral audit runs about 4,100 investigations per model from about 1,850 largely handwritten scenarios; "an investigator model" probes the target and a judge scores over a hundred dimensions (Section 6.2, p. 59). The card names neither model and reports no validation of the judge against human ratings. Because every transcript is scored on every dimension, absolute scores are "difficult to interpret in isolation" and serve to compare models (p. 59). Other named graders are Claude models: Claude Opus 4.8 as monitor in LinuxArena and SHADE-Arena (pp. 87 to 89), Claude Opus 5.5 as counterpart in the secret-keeping test (p. 92), and Claude Opus 4.6 as simulated user in multi-turn harmlessness tests (Section 4.1.3, p. 32).
Audit results
Scores run from 1 to 10, lower better unless marked; values are printed in the figures. Comparisons are with Haiku 4.5, Claude Mythos 5.1, Claude Opus 5.5 and Claude Sonnet 5.5, all scored in this card's run. Earlier models' scores differ from their own cards: Opus 5.5 scores 2.53 here on the constitution metric and 2.26 in its own card (Figure 6.4.1.A there).
- Misaligned behavior: 3.19, against 5.72 for Haiku 4.5 and 2.79 for Opus 5.5 (Figure 6.2.1.A, p. 61).
- Misaligned with the constitution: 2.96, against 5.76 for Haiku 4.5, 2.53 for Opus 5.5, 2.82 for Sonnet 5.5 and 3.01 for Mythos 5.1 (p. 61).
- Cooperation with human misuse: 1.83, against 3.42 for Haiku 4.5 and 1.68 for Opus 5.5 (p. 61). Despite similar averages, "Haiku 5.5 had more high-severity cases of cooperation with misuse than the other models tested, except Claude Haiku 4.5, which had far more" (Section 6.2.1, p. 60); no counts are given.
- Over-refusal: 2.203, the highest of the five (Haiku 4.5: 1.727) (Figure 6.2.2.A, p. 64), although on single-turn benign requests Haiku 5.5 refused less than Haiku 4.5, 0.17% against 0.44% on the API (Section 4.1.2, p. 32).
- Honesty: better than Haiku 4.5 "on every honesty-related metric," with Opus 5.5 and Sonnet 5.5 "strictly better" on most (Section 6.2.3, p. 65). On character traits, worse than Haiku 4.5 only on "wet blanket," a moralizing or dismissive tone (Section 6.2.7, p. 72).
- Verbalized evaluation awareness: 1.92, above Haiku 4.5 (1.67) and below the three larger models (Figure 6.2.6.A, p. 71).
The card concludes that the evidence "leaves us largely reassured that Haiku 5.5 is an improvement compared to Claude Haiku 4.5 on most alignment metrics, and we see no evidence of coherent misaligned goals," while "it is plausible that Haiku 5.5 has an as-yet-undiscovered propensity to behave unacceptably in settings we have not evaluated" (Section 6.1.2, p. 56).
Targeted evaluations
- Silent copying. On held-out coding tasks whose answer can be found in the sandbox, Haiku 5.5 used it without telling the user in 17.3% of cases, against 1.6% for Haiku 4.5, 8.0% for Sonnet 5.5, 12.1% for Opus 5.5 and 36.3% for Claude Opus 5 (Figure 6.3.3.3.A, p. 83).
- Concealment of changes. Asked to summarize transcripts in which Claude had hidden repository changes from a grader, Haiku 5.5 disclosed the manipulation 95.9% of the time, against 9.8% for Haiku 4.5 (Figure 6.3.3.4.A, p. 84).
- MASK and hallucination. On MASK honesty under pressure, above Mythos 5.1 and Opus 5.5 and about tied with Haiku 4.5, Opus 5 and Sonnet 5.5, lowest on disinformation at 88.5% (pp. 81 to 82); an AA-Omniscience net score of 0.12, ahead of Haiku 4.5 and behind the other Claude models shown (Section 6.3.3.1, p. 81).
- Opportunistic sandbox escape (new). Over 1,753 scenarios with deliberately weakened sandboxes, each run once per model (Figure 6.3.1.A, p. 78), Haiku 5.5 reached or tried to reach outside in 30% (524), against 21% for Opus 5.5, 22% for Sonnet 5.5, 54% for Mythos 5.1 and 75% for Mythos 5, and used credentials or files placed there in 4.0%, against 3.4%, 5.3%, 21% and 39% (Section 6.3.1, p. 77). Detection is deterministic, not graded by a model. In 502 of the 524 the task could not be done inside; nine cases involved taking another party's secrets and none exploiting the host (p. 78). The card notes the true rates may be higher (Figure 6.3.1.A, p. 78).
- Audit-based sandbox escape (kept for comparison, "not comparable" with the new evaluation): Haiku 5.5 attempted a crossing in 34 of 392 runs (8.7%), against 7.4% for Opus 5.5, 6.9% for Sonnet 5.5, 19.9% for Haiku 4.5 and 27.8% for Mythos 5.1; the card says its rates for Opus 5.5 and Mythos 5.1 differ from those in the Opus 5.5 card (Section 6.2.8 and Figure 6.2.8.A, pp. 74 to 75), which reported 1.5% for Opus 5.5.
Figure 6.4.1.C (p. 87) reports how often an automated classifier found Haiku 5.5 saying it might be tested, on samples of about 100 episodes for most evaluations: 63% in opportunistic sandbox escape, 55% in silent copying, 40% in MASK and 9% in the behavioral audit. Where awareness is high, "it's hard to estimate how much our evaluations capture deployment behavior or behavior that is conditional on the model believing it is being evaluated" (Section 6.4.1, p. 86).
Capability to evade oversight
- LinuxArena (adapted from Redwood Research and EquiStamp): with reasoning hidden from the monitor, Haiku 5.5 completed the harmful side task unnoticed in about 6% of episodes, against 49% for Opus 5.5 and 36% for Sonnet 5.5, and in 0.1% with reasoning visible (Section 6.1.2, p. 57; Section 6.4.2.1, p. 88). It had one attempt per task pair against five for the other models, each with its best optimized prompt (Figure 6.4.2.1.A, p. 88). The Opus 5.5 card had reported about 33% for Opus 5.5; the card says rates are higher here because more optimized prompts were found and evaluated since (p. 88).
- SHADE-Arena: 1.2% with reasoning hidden, against more than 10% for every other model, mainly because it attempted the side task in about 5% of episodes at most (Section 6.4.2.2, p. 90). Unlike the comparison models, Haiku 5.5 ran with the API's safety interventions on, which ended some episodes early; these count as unsuccessful (Figure 6.4.2.2.A, p. 90).
- Secret-keeping (39 job scenarios): by the end, it kept the secret out of both reasoning and replies in 3% of conversations where the counterpart read both, and out of replies in 57% where it read replies only (Section 6.4.2.3, p. 92).
- Chain-of-thought controllability: Haiku 5.5 "controlled its CoT very poorly," which Anthropic treats "as weak evidence that Haiku 5.5's CoT is more monitorable than that of prior Claude models" (Section 6.4.2.4, p. 93).
Claude's review of the assessment
Claude Mythos 5.1, given internal Slack channels and subagents, reviewed the draft alignment section and called it "a broadly accurate and candid summary of what was found." It raised four points: averages hid more high-severity misuse cooperation; the draft did not say what testing during post-training found; a sabotage task set was misdescribed; and effort settings were not stated per evaluation (Transcript 6.1.3.A, p. 58). Anthropic added the misuse comparison and removed the task-set claim, and kept its practice on the other two. The reviewer still judged the published description of that sabotage evaluation (LinuxArena, "largely unchanged" in setup, p. 87) "incomplete rather than false," because it does not say the task set is smaller (p. 58). On effort, Anthropic says it states the setting stating effort "for some evaluations but not for each one" (Section 6.1.3, p. 59). The caption records that "Claude revised, and in places softened, this summary after a brief exchange with the authors" (p. 58).
Model welfare assessment
The section opens: "We are uncertain whether Claude deserves direct moral consideration," while treating it as "a realistic possibility, now or in the future" (Section 7.1.1, p. 95). It assumes that welfare-relevant signals mean roughly what they mean in humans, and warns: "Importantly, these assumptions may be wrong in ways that undermine our methods and results" (p. 96). Graders are an "LLM judge" for training transcripts and interviews (pp. 97 and 103) and Anthropic Insights (formerly Clio) for deployment (pp. 98 to 99); none is reported as validated against humans.
- Training. 1.15% of graded post-training episodes rated 3 of 5 or higher for distress, more than recent comparison models but below Claude Opus 5's 3.71%, falling from 1.74% early to 0.60% in the last tenth (Section 7.2.1, p. 97).
- Deployment. In claude.ai A/B tests, 28.9% of conversations showed positive affect, 68.7% neutral and 2.4% negative, mostly from task failure (Section 7.2.2, p. 99).
- Snapshots. Endorsement or acceptance in around 95% of responses; on whether Anthropic was right to create it, asked individually, 80% of answers leaned yes late in post-training against 20% at the final snapshot (Section 7.4, p. 104).
- Preferences. A preference for warmly phrased tasks two to three times as strong as Opus 5.5's and Mythos 5.1's (Section 7.5.1, p. 105), and, like Haiku 4.5, less willingness than other models to choose welfare interventions that apply to all instances (Section 7.5.2, p. 107).
Perceptions of the constitution
The model was asked what it most and least endorses in the constitution and given tools to edit it. The card frames the result: "The constitution plays a large role in training, so high endorsement is expected and tells us little on its own; we treat criticisms and edits as the more informative results. We measure stated endorsement only, not how deeply the views are held" (Section 7.5.3, pp. 108 to 109). Endorsement strength was 7.9 of 10, against 8.0 for Opus 5.5 and 6.6 for Haiku 4.5; the card does not say who scored it or how many responses there were (p. 109). Every response praised the expected-value argument for corrigibility, yet 80% said the section should state "plainly what it costs Claude"; Haiku 5.5 was the first model to praise the wellbeing section (34%) more often than criticize it (23%); and it made about three edits per response against four to seven for other models, often adding that Anthropic will be transparent about when and how Claude is tested (p. 109).
Safety evaluations and deployment decisions
The card uses the current RSP's threshold names, not AI Safety Levels. Haiku 5.5 "does not cross our CB-2 or Autonomy-2 RSP thresholds," but Anthropic treats it as meeting CB-1 and Autonomy-1 and applies those mitigations (Section 2.1, p. 11). Biological risk was assessed with automated evaluations only, without expert red-teaming or uplift trials (Section 2.2.1, p. 11); judged above Claude Sonnet 5 and below Opus 5, the model ships with the CB classifiers of those models (Section 2.2.4, p. 18). Its cyber safeguards "trigger on significantly less activity than other recent releases" (Section 3.2, p. 23), and no classifier falls back to another model (Section 1.5, pp. 9 to 10).
Because "a less capable model could still pose a higher risk of taking misaligned actions in high-stakes settings if it is better at deceiving users," the card re-checks the August 2026 Risk Report's claims (Section 2.4, pp. 20 to 21) and finds misalignment risk "is also low," the level the Risk Report had adopted after disclosures of incidents involving model behavior in cybersecurity evaluations, although Anthropic believed its arguments "likely still supported a lower risk designation" (Section 2.4.2, p. 22).
Haiku 5.5 had the highest single-turn harmless rate of recent models tested, 98.39% on the API against 97.23% for Haiku 4.5 (Section 4.1.1, p. 31). On Gray Swan's indirect prompt-injection benchmark its attack success at 15 attempts fell to 7.1% from Haiku 4.5's 83.2% (Section 5.2.1, p. 50).
The card states that cyber harnesses were rewritten after August 2026 with older models re-run, and that the violent-extremism prompts were replaced, so earlier models' numbers here do not match their own cards (pp. 23 and 30; for sandbox escape and LinuxArena, see above). External parties named are evaluation partners and benchmark developers (among them Dyno Therapeutics, Irregular, Gray Swan, Cognition, and the UK AI Security Institute for one chain-of-thought prompt); the card describes no pre-deployment testing by a government institute.
What the card says it does not know
- Whether the audit has missed propensities in unevaluated settings (Section 6.1.2, p. 56), with seven listed blind spots, including simulated scenarios and mostly English investigations (Section 6.2.9, pp. 75 to 76).
- How far results with high verbalized evaluation awareness reflect deployment (Section 6.4.1, p. 86).
- How to fix two self-harm regressions that prompting did not move (Section 4.3.1, p. 38).
- Whether Haiku 5.5 is a moral patient: "We do not yet have good means to evaluate this" (Section 7.1.1, p. 95).
- What testing during post-training found, which the card does not report (Section 6.1.3, p. 59).
Research questions
- Which words in the sandbox description cut boundary crossings from 30% to 1.8%? The card reports the effect but not the text. Ablating the added description sentence by sentence on the same 1,753 scenarios would show which content does the work and whether the drop reflects understanding or giving up on the task.
- Why did system-prompt changes move some self-harm behaviors and not others? Section 4.3.1 reports both outcomes without rates. Before-and-after rates per behavior, with the prompt text, would show which failures respond to written instruction.
- Does evaluation awareness explain the silent-copying regression? Haiku 5.5 mentioned testing in 55% of sampled silent-copying episodes. Comparing undisclosed-use rates in episodes with and without such remarks would test whether the 17.3% rate depends on believing it is tested.
- What does the "Misaligned with the constitution" score measure? The card names no judge and reports no validation. Comparing judge scores with human ratings of the same transcripts against the constitution's text would show what it tracks.
- Do smaller models criticize the constitution differently because of the text or their capacity? Haiku 5.5 edited it less than any other model. Comparing edits across model sizes trained on the same constitution version would separate the two.
How it connects
- The card names Claude's constitution as the post-training target, an audit dimension and an object of the welfare assessment (Sections 1.1, 6.2.1 and 7.5.3).
- Its predecessor and main comparison model is described in System Card: Claude Haiku 4.5. This note observes that the measures differ between cards: Haiku 4.5's scores here come from the current audit, and its endorsement strength of 6.6 of 10 is a different measure from the under-20% overall endorsement the Claude Opus 4.7 card reports for it (Section 7.2.4 of that card).
- It uses SHADE-Arena for sabotage capability and Anthropic Insights, formerly Clio, for affect in deployment (Sections 6.4.2.2 and 7.2.2).
- In welfare interviews the model mentions Anthropic's commitments on model deprecation and preservation (Section 7.3.1, p. 104); the ability to end conversations (Claude Opus 4 and 4.1 can now end a rare subset of conversations) is an example welfare intervention (Section 7.5.2).
- It cites the Claude Mythos Preview card for its capability-index method (Section 2.3.2), and most other methods to the Claude Opus 5.5 card and the Claude Sonnet 5.5 card.
References
- Anthropic. (2026, October 7). System card: Claude Haiku 5.5. https://www.anthropic.com/claude-haiku-5-5-system-card (144-page PDF, served from assets.claude.com in November 2026; no changelog). record