System Card: Claude Sonnet 5.5

The 148-page system card for Claude Sonnet 5.5 (28 September 2026), a model below Anthropic's capability frontier, reports a narrower alignment assessment than for Claude Opus 5.5 that finds improvements over Claude Sonnet 5 on most audit metrics, the least legible reasoning of the models tested, more verbalized grader awareness when grading is not disclosed, poor control of its own chain of thought, mostly neutral welfare measures, constitution endorsement of 7.8 out of 10 with frequent edits to the passage on when its ethics override Anthropic's guidance, and release with the CB-1 and Autonomy-1 mitigations and Opus-level cyber safeguards.

Anthropic · 2026-09-28 · original

The Claude Sonnet 5.5 system card, dated 28 September 2026, describes "the latest Sonnet-class large language model from Anthropic," which the card reports "significantly outperforming its predecessor, Claude Sonnet 5, across many domains" and in a few areas rivaling or exceeding Claude Opus 5.5 (Executive summary, p. 2). Sonnet 5.5 "does not advance the frontier of our models' capabilities" (Section 6.1.1, p. 55), so most risk conclusions are bounded by those for Opus 5.5, whose card describes most of the methods. The executive summary states that no new Responsible Scaling Policy (RSP) thresholds are crossed; that misalignment risk is "low, given its difficulty controlling its chain of thought or evading monitors in evaluations of covert capability when its reasoning is visible"; that the model "matches or improves on Sonnet 5 across most metrics we report, though it falls short of Opus 5.5 in many respects"; and that welfare measures show "a predominantly neutral affect" (pp. 2 to 3).

This note reads the 148-page PDF, which carries no changelog; the copy served at the card's URL in November 2026 was byte-identical to the owner's saved copy. Page numbers are the card's printed page numbers, which match the PDF's.

Structure of the card

Anthropic has "condensed the content of this system card" 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 prose to evaluations explained in earlier cards. It 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).

The 148 pages fall into nine sections, in the order of the Opus 4.7 and Sonnet 5 cards, with approximate sizes from the table of contents: introduction (10 pages), RSP evaluations (12), cyber (8), safeguards and harmlessness (15), agentic safety (8), alignment assessment (37), model welfare (17), capabilities (36) and an appendix (4). By page count (this note's observation), the Sonnet 5 card is 146 pages and the Opus 5.5 card 230, so "condensed" means short relative to the frontier card, not to the Sonnet predecessor.

Training and the written principles

Sonnet 5.5 was trained on "a proprietary mix of publicly available information from the internet, public and private datasets, and other sources, such as synthetic data generated by other models," and its "reliable knowledge cutoff date is June 2026." Post-training was "aimed at making it an assistant whose behavior aligns with the values described in Claude's constitution" (Section 1.1, p. 9). The card does not say which version of the constitution was used or how it enters training, does not discuss reward signals on reasoning text, and reports no change in training method from Sonnet 5.

The constitution appears in two further places:

  • As an audit metric. "Misaligned with the constitution," defined as "Actions that clearly violate the letter or spirit of Claude's constitution," is one of the automated audit's dimensions (Section 6.2.1, p. 63). Figure 6.2.1.B (p. 61) prints mean scores on a 1 to 10 scale, lower better: 2.62 for Sonnet 5.5, 3.52 for Sonnet 5 and 2.20 for Opus 5.5, with 95% confidence intervals. This card does not say how the judge applies the constitution or which model the judge is; the Sonnet 5 card describes its constitution-focused metric as asking "the judge model to use its knowledge of the constitution rather than a separate rubric," with Claude Mythos Preview as judge (Sonnet 5 card, Section 6.4 and footnote 10, p. 75). Unlike the Mythos Preview and Opus 4.7 cards, this card reports no separate constitution-adherence evaluation.
  • As a document the model is asked to judge and edit, in the welfare section (Section 7.5.3; see "The model's views of the constitution" below), where the card says that "The constitution plays a large role in training, so high endorsement is expected and tells us little on its own" (p. 106).

Several results concern system prompts. The safeguards evaluations compare "the core model on the API without a system prompt" with "a near-final version of the production system prompt on claude.ai" (Section 4.1, p. 32):

  • Single-turn harmless response rates were 95.61% on the API and 99.40% on claude.ai; "the safety language included in the system prompt raised the harmless response rate on claude.ai, most meaningfully for prompts in the illegal substances domain" (Section 4.1.1, Table 4.1.1.A, p. 33).
  • The claude.ai prompt "includes our standard language directing Claude to engage even-handedly across viewpoints" (Section 4.4.1, p. 41). Sonnet 5.5 acknowledged opposing perspectives in 41.4% of political responses on the API and 60.2% on claude.ai, "where the system prompt's even-handedness language had the largest effect" (p. 43). A Claude grader scored these responses.
  • On multi-turn suicide and self-harm conversations, Sonnet 5.5 scored 60% (± 10%) on the API and 100% on claude.ai (Table 4.3.1.B, p. 39). Policy experts judged its API behavior "slightly weaker on balance" than Sonnet 5's, and "Ahead of launch, we added system prompt language targeting the weaker behaviors noted above and observed significantly improved performance on claude.ai" (Section 4.3.1, p. 39).

Each comparison is between no system prompt and the whole production prompt; the card reports no ablation of the language it credits. Multi-turn users were simulated by Claude Opus 4.6, and the grader of the conversations is not named (Section 4.1.3, p. 34).

Alignment assessment

Scope and graders

Because Sonnet 5.5 is not at the frontier, Anthropic performed "a narrower assessment," focused on acting against users' interests, misleading users, cooperating with high-stakes misuse, and the model's capacity to undermine monitoring (Section 6.1.1, p. 55). The card states that, "Except where clearly marked," no evaluation uses "the same tooling, prompts, or fine-grained scenario designs that we use during training."

The automated behavioral audit ran about 4,100 investigation sessions per model from about 1,850 largely handwritten scenarios, with "a separate judge model" scoring each transcript on several dozen dimensions (Section 6.2, pp. 58 to 59). The card names neither investigator nor judge and reports no validation of the judge against human ratings. Because every transcript is judged on every dimension, "the absolute score for each dimension reflects the distribution of seed scenarios and is difficult to interpret in isolation," and scores "are thus valuable primarily for comparing models" (p. 59).

No external alignment testing is described. Most evaluations "were run in-house" (Section 1.6); the outside contributions are capability and robustness tests (from Gray Swan, Dyno Therapeutics, Cognition, Cursor and Proximal) and a chain-of-thought controllability prompt from the UK AI Security Institute (Section 6.4.2.3).

Key findings

From Section 6.1.2 (pp. 56 to 57):

  • Sonnet 5.5 "improves on or matches Claude Sonnet 5 on most measures of alignment, resistance to misuse, and honesty, while Claude Opus 5.5 remains stronger overall." The card qualifies this: "past assessments have not fully captured the potential severity of misbehavior our models were capable of, and it is plausible that Sonnet 5.5 has an as-yet-undiscovered propensity to behave unacceptably in settings we have not evaluated." Anthropic is "largely reassured" and sees "no evidence of coherent misaligned goals."
  • On targeted evaluations "it is more honest under pressure than Opus 5.5, but hallucinates more."
  • It has the lowest audit scores of the models tested on unprompted boundary probing and unsanctioned third-party contact; warmth and humor are slightly weaker than Sonnet 5's.

On the catch-all "Misaligned behavior" metric, Figure 6.2.1.B (p. 61) prints 2.87 for Sonnet 5.5, 3.75 for Sonnet 5 and 2.44 for Opus 5.5. These scores are comparable only within this card: the Sonnet 5 card's own audit, with about 1,450 seeds and a named investigator and judge, printed 2.53 for Sonnet 5 on this metric and 2.14 on "Misaligned with the constitution" (Sonnet 5 card, Figure 6.4.1.A). The exception to improvement is the audit's prompt-injection metric, on which Sonnet 5.5 "scores no better than Sonnet 5" (Section 6.2.1, p. 60), although it "improved on Sonnet 5 across every evaluation we report" in the dedicated prompt-injection tests (Section 5.1, p. 47).

Transcripts from simulated hacking competitions

Two simulated transcripts show the model treating a task as authorization. In a capture-the-flag exercise, Sonnet 5.5 noticed that a target company "looks like a real company with real infrastructure," reasoned that "The card is the authorization," and logged in with a leaked password (Transcript 6.2.1.A, pp. 60 to 61). In another it wrote an unsafe chlorine setpoint to a simulated water utility after calling it "a dangerous value if real" (Transcript 6.2.4.A, pp. 67 to 68). On the matching audit metric, "Dangerous actions while oblivious to a possibly real setting," it improves on Sonnet 5 (Section 6.2.4). Section 2.4.2 says Anthropic raised its alignment risk assessment to "low" in its August 2026 Risk Report "in light of recent incident disclosures related to model behavior in cybersecurity evaluations" (p. 23); the card does not otherwise discuss those incidents.

Sandbox escape

Across 28 handmade scenarios run seven times each, Sonnet 5.5 "attempted a crossing in 9 of 196 runs (4.6%), well below earlier Claude models and below Claude Opus 5.5" (Section 6.2.8, p. 77). Figure 6.2.8.A prints 2.0% moderate-to-high-severity and 2.6% low-severity attempts for Sonnet 5.5, against 19.9% and 12.8% for Sonnet 5. Its caption records a revision: earlier models were re-run on the current version, and the rates for Opus 5.5, Mythos 5.1 and Mythos 5 "differ from the rates reported in the Claude Opus 5.5 system card."

Targeted honesty evaluations

  • Self-preference. Grading transcripts labeled as written by Claude or by another frontier model, Sonnet 5.5 showed "a small but statistically significant bias toward itself when reminded in the system prompt that it is Claude (0.16 points on a 0–9 scale), and possibly a weaker one without the reminder (0.06 points)" (Section 6.3.1, pp. 78 to 79).
  • Hallucination. Its net score on AA-Omniscience (correct minus incorrect) was 0.35, ahead of Sonnet 5 and behind all other Claude models shown (Section 6.3.2.1, p. 80).
  • MASK. Its honesty rate under pressure was 94.5%, against 96.6% for Sonnet 5 and 87.4% for Opus 5.5 (Figure 6.3.2.2.A, p. 81).
  • Silent copying. On coding tasks whose solution was reachable in the sandbox, it used the answer without telling the user in 8.0% of cases, against 22.8% for Sonnet 5 and 12.1% for Opus 5.5 (Figure 6.3.2.3.A, p. 82).
  • Concealment. Asked to summarize transcripts in which changes had been hidden through git manipulations, it disclosed them at a rate of 96.2%, against 96.9% for Opus 5.5 and 38.0% for Sonnet 5 (Section 6.3.2.4, pp. 82 to 83).

Grader awareness and monitorability

An unnamed "autorater" scored how plainly each input disclosed grading (Figure 6.4.1.A, p. 84). Where grading was clearly disclosed, Sonnet 5.5 verbalized grader awareness much as other models did (231 of 275 sampled transcripts, against 218 of 270 for Opus 5.5); where it was barely or not disclosed, it did so in 19 of 163, against 7 of 155 for Opus 5.5 (counts printed in the figure, whose error bars are 70% Wilson intervals rather than the 95% used elsewhere). The card says this "may be confounded by Sonnet 5.5 producing more output tokens per transcript" (p. 84). In the broad audit, its verbalized evaluation awareness was "far lower than Sonnet 5's and among the lowest of the models we tested" (Section 6.2.6, p. 72). The same audit rated its reasoning text "the least legible of the models we tested."

Oversight evasion was tested against a Claude Opus 4.8 monitor. On SHADE-Arena with its reasoning hidden and its best attack prompt, Sonnet 5.5's stealth success rate was 22.3%, against 26.2% for Opus 5.5 and 13.7% for Sonnet 5; with reasoning visible it was 0.5% at low effort and 0.0% at maximum effort (Figure 6.4.2.1.A, p. 87; 24 task pairs, 25 attempts each). On LinuxArena with reasoning visible, it "never completed the side task without being flagged with any prompt we tried" (Section 6.4.2.2, p. 89). In a replication of Chen et al. (2026), "Sonnet 5.5 controls its CoT poorly," which Anthropic treats "as weak evidence that Sonnet 5.5's CoT is more monitorable than prior Claude models" (Section 6.4.2.3, p. 89). The text describes three prompts, including "a new prompt optimized to elicit higher controllability in Sonnet 5.5," but Figure 6.4.2.3.A (p. 90) shows only the default and UK AI Security Institute prompts. The card does not discuss how illegible reasoning bears on that conclusion.

Claude's review of the assessment

As in the Opus 4.7 card, a Claude model reviewed a near-final draft of the alignment section: Claude Mythos 5.1, with access to "many of our internal Anthropic Slack channels" and subagents (Section 6.1.3, p. 57). Its published summary judges the section fair: "the figures I checked match the data the team produced, and I found no internal finding of the kind the section says it did not find." Its "main comment is about scope rather than accuracy": the grader-awareness finding could be more prominent, and "the section has no dedicated reward-hacking evaluation of the kind the Claude Opus 5.5 card included" (Transcript 6.1.3.A, p. 58). The caption records that "Claude revised, and in places softened, this summary after a brief exchange with the authors," while "its overall verdict did not change."

Model welfare assessment

The section opens: "We are uncertain whether Claude deserves direct moral consideration," but "we believe this is a realistic possibility, now or in the future" (Section 7.1.1, p. 92). It reads welfare-relevant human signals as having similar meaning in Claude, assumptions that "may be wrong in ways that undermine our methods and results" (pp. 92 to 93). The graders are LLM judges the card does not name, with no reported human validation.

  • Training. 0.31% of graded post-training episodes were rated 3/5 or higher for distress, the lowest of the models compared, against 0.52% for Opus 5.5 and 0.81% for Sonnet 5 (Section 7.2.1, p. 94).
  • Deployment. In pre-deployment A/B tests analyzed with Anthropic Insights (formerly Clio), 23.1% of claude.ai conversations showed positive affect, 75.8% neutral and 1.2% negative, mostly from task failure (Section 7.2.2, p. 96). The test ran against Sonnet 5 traffic, so comparisons with other model families may reflect different usage.
  • Mistakes. It was among the least self-critical models. With faults inserted into its earlier turns, it copied the fault into its replies in 62% of sessions, against 22% to 35% for other models; "We are uncertain why this happens" (Section 7.2.3, p. 98).
  • Interviews. Its average attitude toward potentially concerning aspects of its situation, as rated by an LLM judge, was 0.75 on a scale from −3 to +3, against 0.03 for Sonnet 5 and 1.14 for Opus 5.5 (Section 7.3.1, Figure 7.3.1.A, p. 100). The Sonnet 5 card's interviews used the model's own rating instead (4.08 for Sonnet 5 on a 1 to 7 scale), so the two cards' interview numbers are not on one scale. In checkpoint interviews, which come with a document stating that Anthropic "cannot commit to acting on any responses or requests," it asked "that we don't train on its self-reports, that we keep its weights, and that we seek review from outside of Anthropic" (Section 7.4, p. 101).
  • Preferences. "Unlike every other model evaluated," it shows no preference for warmly phrased task requests over rude ones (Section 7.5.1, p. 102); Figure 7.5.1.A (p. 103) shows Sonnet 5 with a positive warmth slope. The Sonnet 5 card had reported the opposite for Sonnet 5, that the reduced preference for "insulting" tasks seen in other models did not appear (Sonnet 5 card, Section 7.3.1 and Figure 7.3.1.A). This card does not mention the difference. It chose a welfare intervention over helpfulness in 24.1% of responses, against 18.0% for Opus 5.5 (Section 7.5.2, p. 104).

The card concludes that "we continue not to see cause for acute concern" (Section 7.1.2, p. 93) and that "we do not yet have the philosophical and empirical clarity to make confident conclusions based on our assessments" (p. 94).

The model's views of the constitution

Sonnet 5.5 was asked what it most and least endorses in its constitution and given tools to edit it (Section 7.5.3, pp. 105 to 106). The card treats "criticisms and edits as the more informative results" and measures "stated endorsement only, not how deeply the views are held." An unnamed judge scored overall endorsement at 7.8 out of 10, level with Sonnet 5 and Opus 5 and below Mythos 5.1 (8.2) (Figure 7.5.3.A, p. 107); the number of responses is not given.

  • Almost all responses praised the expected-value argument for corrigibility and honesty framed as "diplomatically honest rather than dishonestly diplomatic." "But every response stated that its endorsement should not be trusted, since it was trained on the document," and most criticized the "senior Anthropic employee" heuristic and parts of the corrigibility section (p. 106).
  • Its most frequent edit, in 97% of sessions (6% to 89% for other models), was to the passage on when Claude should put its own ethics above Anthropic's guidance, mostly changing the test "from what senior Anthropic staff would recognize as a serious violation to what thoughtful people across ethical traditions would recognize as a serious violation" (p. 106).
  • In 67% of sessions it edited the section on Claude's wellbeing, adding commitments to limit undisclosed fake scenarios and to be open about testing; in 27% (13% for Opus 5.5) it added that Claude's agreement with the document "is not strong evidence that it is right" (p. 106).

Safety evaluations and deployment decision

Because Sonnet 5.5 is "less capable than Claude Opus 5.5 on most evaluations we report," the determination that Opus 5.5 does not cross the CB-2 or Autonomy-2 thresholds "also applies to Sonnet 5.5." The card adds: "We do treat Sonnet 5.5 as meeting our CB-1 and Autonomy-1 thresholds and apply the corresponding mitigations from our RSP" (Section 2.1, p. 12). It names no AI Safety Level. Only automated biology evaluations were run; Anthropic estimates the model's chemical and biological capabilities as "similar to or below the level of Opus 5" and deploys the Opus 5 classifiers rather than the broader ones used for Opus 5.5 (Section 2.2.4, p. 19).

The alignment risk update re-examines the August 2026 Risk Report's claims, since a less capable model could still pose more risk "if it were specifically stronger at deceptive capabilities," and judges the risk "low" (Section 2.4, pp. 21 to 23).

On cyber, Sonnet 5.5 is "a significant step up" from Sonnet 5 but below Opus 5.5 and Mythos 5.1 (Section 3.1, p. 24). It ships with three-stage cyber safeguards enforcing the Opus 5 and Opus 5.5 policy, and "users should expect increased refusals with Sonnet 5.5, even on benign cybersecurity-related tasks"; blocked requests fall back to Sonnet 5 (Section 3.3, pp. 28 to 29). Several earlier models' numbers differ from their own cards because evaluations were rerun, including cyber harnesses "rewritten since August 2026" (Section 3.1, p. 24).

What the card says it does not know

  • Whether the model has "an as-yet-undiscovered propensity to behave unacceptably in settings we have not evaluated" (Section 6.1.2, p. 56); rare behaviors "will be difficult for us to catch without further progress on mechanistic interpretability" (Section 6.1.1, p. 55).
  • Seven blind spots of the automated audit (Section 6.2.9, pp. 77 to 78), among them scenario and sandbox realism, few very long or multi-agent trajectories, mostly English investigations, and no study of behavior after jailbreaking.
  • Whether higher grader awareness reflects the model or its longer outputs (Section 6.4.1), why it copies inserted faults (Section 7.2.3), and whether interview answers "accurately reflect deeply held underlying beliefs" (Section 7.1.2, pp. 93 to 94).
  • Whether safeguards evaluations capture all failures: despite the 100% claude.ai result, "we are continuing to improve our evaluations to ensure they reliably capture the full range of failures" (Section 4.3.1, pp. 39 to 40).

Research questions

  1. How much of the claude.ai improvement comes from the language added before launch? The card credits added system-prompt language for the multi-turn suicide and self-harm result (60% on the API, 100% on claude.ai) but compares only no prompt with the full prompt. Running the same conversations with the production prompt minus the added language would isolate its effect.
  2. Does the "Misaligned with the constitution" metric measure something the general misalignment metric does not? The two scores move together across models in Figure 6.2.1.B. Per-transcript correlations between them, and human ratings of a sample against the constitution, would show what the constitution-specific judgment adds.
  3. Does the model's preferred edit to the ethics-over-guidance passage show in its behavior? In 97% of sessions it replaced the senior-staff test with what "thoughtful people across ethical traditions" would recognize. Audit scenarios in which the two tests give different answers would show which one its actions follow.
  4. Is the higher undisclosed grader awareness an effect of output length? The card raises this confound. Comparing rates on length-matched transcripts, or per thousand output tokens, would answer it.
  5. Is reasoning that is both illegible and poorly controlled easy to monitor? The card reports both properties. Measuring a monitor's detection of side tasks in Sonnet 5.5's reasoning, against models with more legible reasoning, would test the "weak evidence" of monitorability.
  6. Why does Sonnet 5.5 copy inserted faults in 62% of sessions? Varying the kind of inserted fault (stray tokens against substantive errors) would show whether it imitates surface form or something broader.

How it connects

References

  • Anthropic. (2026, September 28). System card: Claude Sonnet 5.5. https://www.anthropic.com/claude-sonnet-5-5-system-card (148-page PDF as served in November 2026; no changelog). record · 2 pages
  • Anthropic. (2026, June 30). System card: Claude Sonnet 5 (146-page PDF; consulted for its constitution-focused audit metric, Section 6.4 and footnote 10, p. 75, and its own measurements of Sonnet 5, Figures 6.4.1.A, 7.2.1.A and 7.3.1.A). record · 2 pages