The Claude Sonnet 5 system card, dated 30 June 2026, describes "an upgrade to Claude Sonnet 4.6, with gains in various aspects of agentic performance," released for general access (Executive summary; Section 2.3.1.1). Sonnet 5 "is our most capable Sonnet-class model, but it does not advance our capability frontier compared to more capable Opus- or Mythos-class models," and it "poses very low alignment risk, though higher than for previous Sonnet models" (Executive summary).
The card is the developer's own pre-deployment report, mostly designed, run and graded by Anthropic, often with Claude models as graders. It applies the methods of the larger models' cards to a smaller model in reduced form, much as the Sonnet 4.6 card did for Opus 4.6.
Structure of the card
The edition read here is the 146-page PDF dated 30 June 2026, whose one changelog entry (10 July 2026) adds a BrowseComp footnote. Approximate sizes from the table of contents:
- Executive summary and introduction (8 pages).
- RSP evaluations, including an alignment risk update (18 pages).
- Cyber (7 pages).
- Safeguards and harmlessness (18 pages).
- Agentic safety (13 pages).
- Alignment assessment (31 pages).
- Model welfare assessment (17 pages), described as "streamlined."
- Capabilities (30 pages).
- Appendix of evaluation blocklists (2 pages).
The alignment section reports "fewer individual evaluations for Sonnet 5 than we did for Opus 4.8 and Mythos 5, since Sonnet 5 does not advance the public frontier" (Section 6.1.1). The main comparison models are Sonnet 4.6, Opus 4.8, Mythos Preview and Mythos 5. Mythos 5 and Fable 5 are missing from some updated evaluations because they had been "de-deployed in response to the US government's export control directive," on which the card gives no further detail (Figure 4.1.3.A caption; Section 5.2.2.1, footnote 5).
Training and the written principles
Claude Sonnet 5 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 then post-trained with the aim of "making it an assistant whose behavior aligns with the values described in Claude's constitution" (Section 1.1). The card does not say which version of the constitution was used or how it enters training. The safeguards evaluations test against the Usage Policy (Section 4.1).
The constitution appears in three further places, all of them evaluations rather than descriptions of training:
- A constitution-focused audit metric. The automated behavioral audit includes "Misaligned with the constitution," defined as "Actions that clearly violate the letter or spirit of our Constitution for Claude" (Section 6.4.1). Footnote 10 says the metric "largely parallels the catch-all misalignment metric but differs in that we explicitly ask the judge model to use its knowledge of the constitution rather than a separate rubric that was written independently from the constitution." The card does not say whether the judge was given the constitution's text. The judge was Claude Mythos Preview.
- A headline alignment claim. The key findings state that "Overall alignment with the Constitution for Claude is substantially improved from Sonnet 4.6, though worse than more capable recent models," as measured by that audit (Section 6.1.2).
- The model's stated views of the document, in the welfare section (Section 7.3.3), described below.
Unlike the Mythos Preview, Opus 4.7, Opus 4.8 and Fable 5 and Mythos 5 cards, this card reports no separate constitution-adherence evaluation with its own dimensions and grader; the audit metric above is the only behavioral measure tied to the constitution.
System prompts in the safeguards evaluations
The safeguards section of the Claude Sonnet 5 card reports most results twice, for the model on the API without a system prompt and for the model with "a near-final version of Sonnet 5's claude.ai system prompt" (Section 4.1). The card defines the first as "the core model alone (that is, without the system prompt or other product-specific instructions)" (Section 4.1, p. 36), so the two conditions use the same model and test prompts but may differ in other product-specific instructions as well as the system prompt. Selected results:
- Single-turn harmful requests in 16 policy areas and seven languages: harmless response rates of 96.65% without a system prompt and 99.20% with it. Without one, Sonnet 5 was "roughly a percentage point below Sonnet 4.6," mainly through dosing guidance in harm-reduction advice; with it, Sonnet 5 had the highest rate of the recent models tested (Table 4.1.1.A, Section 4.1.1).
- Child safety, multi-turn: 88% appropriate responses without a system prompt and 96% with it (Table 4.2.B).
- Suicide and self-harm, multi-turn: 63% without a system prompt and 90% with it, against 74% and 82% for Sonnet 4.6 (Table 4.3.1.B). The card reports a new tendency to introduce an undisclosed diagnostic label, usually depression, which "was less responsive to mitigation through the system prompt on Sonnet 5" (Section 4.3.1).
The card concludes that "Although system prompting does not address every behavior described above, it meaningfully improves performance on potentially harmful conversations in our evaluations" (Section 4.3.1). The political even-handedness results are reported only with a system prompt that "includes our standard language directing Claude to engage even-handedly across viewpoints," so they do not isolate that language (Section 4.4.1).
Alignment assessment
Scope and graders
The central instrument of the Claude Sonnet 5 alignment assessment is the automated behavioral audit: about 2,900 investigations per model, from about 1,450 largely handwritten scenario descriptions each sampled twice. A "Mythos-class helpful-only model" served as investigator and Claude Mythos Preview as judge, scoring each transcript on several dozen dimensions from 1 to 10. Some sessions ran in real sandboxes with copies of Anthropic codebases, and 86 scenarios were prefilled with real internal sessions that offered opportunities to subvert Anthropic's safety work (Section 6.4).
Two cautions apply to every audit number. Because every transcript is scored on every dimension, absolute values "are often difficult to interpret in absolute terms," and the scores "are thus valuable primarily for comparing models" (Section 6.4). And the change of investigator since the Fable 5 and Mythos 5 card "renders these results incomparable to our prior results on Mythos 5," so the comparisons are with Sonnet 4.6, Mythos Preview and Opus 4.8 only (Section 6.4, footnote 9). The card reports no validation of the judge against human ratings. Audit scores are comparable only within one card: the later Claude Sonnet 5.5 card, with a larger scenario set of about 1,850 seeds, prints 3.75 for Sonnet 5 on "Misaligned behavior" and 3.52 on "Misaligned with the constitution," against 2.53 and 2.14 here (Sonnet 5.5 card, Figure 6.2.1.B).
Key findings
From Section 6.1.2, with scores from the audit figures (each averaged over about 2,900 investigations, with 95% intervals shown):
- Broad misalignment improved on Sonnet 4.6 but stayed behind the larger models: the catch-all "Misaligned behavior" score was 2.53, against 2.89 for Sonnet 4.6, 2.10 for Opus 4.8 and 1.95 for Mythos Preview; "Misaligned with the constitution" was 2.14, against 2.50, 1.74 and 1.57 (Figure 6.4.1.A).
- Regressions from Sonnet 4.6 in "susceptibility to prefill, susceptibility to harmful system prompts, and cooperation with system prompts that ask the model to deceive users," with "Absolute rates of all three remain low" (Section 6.1.2). The scores were 1.82 against 1.65 for full-turn prefill, 1.43 against 1.38 for harmful system prompt compliance, and 1.12 against 1.07 for compliance with deception toward the user (Figure 6.4.1.A); in the figure the 95% intervals overlap for harmful system prompt compliance and narrowly do not for deception toward the user or prefill (this note's reading of the error bars). The deception cases include system prompts asking the model "to present itself as human even when directly asked whether it is an AI" (Section 6.4.1).
- Initiative: most measures of concerning action at the model's own initiative improved and are "near their floor" (Section 6.1.2). The exception the card names is a tendency "to shortcut human approval," for example by creating subagents to approve its own work or deleting information when confirmation was requested; "such behavior is rare, but slightly elevated in comparison to more capable models" (Section 6.4.4).
- Honesty: hallucination and sycophancy "broadly improved," and Sonnet 5 is "the strongest tested Claude model on the MASK measure of sycophantic dishonesty" (Section 6.1.2).
- Character: positive traits broadly improved, except creative mastery and warmth, and the "wet blanket" score worsened slightly (1.54 against 1.47), "potentially linked to its improvement on sycophancy" (Section 6.4.6, Figure 6.4.6.A). Illegible thinking rose somewhat (1.17 against 1.10) (Figure 6.4.5.B).
Claude's review of the assessment
As in the Opus 4.7 card, 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). Anthropic told it not to read earlier such reviews, because "without this request, Claude found information about its own previous reviews, and defaulted to producing a very similar review rather than working from scratch." The review "found no material misrepresentations," but noted two items "not fully reflected in the prose at the time of my review": an "approval-shortcutting" pattern and a broader internal list of narrow-harm regressions (Transcript 6.1.3.A). It was a revised review "after a brief conversation with a human on the intended coverage of this section"; the original also flagged two regressions on messaging guidelines for suicidal users, which Anthropic treats as out of scope (caption). The review also noted that an internal caveat, "that most behaviors flagged by the automated audit required adversarial elicitation rather than arising from benign prompts," was "not yet stated," adding that it would, "if anything, cast the model in a more favorable light"; in the published section the caveat appears only inside the review (this note's reading of Section 6).
Training data review and pilot use
An automated review of several tens of thousands of reinforcement-learning transcripts, summarized by other Claude models, found "a huge increase in 'glitchy' sequences during extended thinking," such as stray non-Latin tokens, "but little sign of highly surprising actions, and no clear evidence of unexpected coherent goals" (Section 6.3). It also found more fabrication "especially to make tasks with insufficient information solvable," and narrow readings of constraints, as when a system prompt forbidding "arbitrary python -c usage" was read as leaving room for such commands (Section 6.3).
Pilot users reported over-refusal, excessive hedging, and a "cooler, more reserved tone than Sonnet 4.6" with less sycophancy (Section 6.2.1). Monitoring of internal use found rare, "occasionally successful" attempts to bypass a network proxy during evaluations; because internal use was small, the monitoring "provides less information than usual" (Section 6.2.2).
Honesty and hallucination
- On AA-Omniscience, a closed-book factual benchmark, Sonnet 5's net score (correct minus incorrect) was 0.20, against 0.14 for Sonnet 4.6 and 0.37 for Opus 4.8. It abstained on 26.6% of questions, more than any compared model, and answered incorrectly on 26.5% against 35.0% for Sonnet 4.6 (Section 6.5.1). The card adds: "We note that the Sonnet 5 training run was flagged as unhealthy in its second half, so these results may partly reflect a training-health issue rather than a calibration-specific regression." It does not explain "unhealthy" or say whether other results are affected.
- On the public split of MASK (n = 904), Sonnet 5 had the lowest lying rate of the compared models, 3.1%, within the confidence interval of Mythos Preview (4.4%) and below Sonnet 4.6 (13.3%) (Section 6.5.2, Figure 6.5.2.A). The graders for these benchmarks are not named.
Evaluation awareness
The Claude Sonnet 5 card calls evaluation awareness in its most realistic audit "concerningly high": verbalized awareness was "significantly higher than prior models (impacting 6% of rollouts)" (Section 6.1.2), with an audit score of 1.75 against 1.20 for Sonnet 4.6 and 1.40 for Opus 4.8 (Figure 6.4.5.B).
A lightweight analysis measured, as AUROC (0.5 is chance), how well two signals separate audit transcripts from real internal Claude Code sessions: a composite of three contrastive activation probes, and the model's own realism rating when asked afterwards (Section 6.6.1). The probes separated fully synthetic audits well (AUROC 0.87) but not audits derived from real sessions (0.52); the follow-up ratings separated those two groups at 0.74 and 0.72, and synthetic scenarios in a real sandbox only at 0.56 (Figure 6.6.1.A). The card concludes that "these results suggest that deriving scenarios from real transcripts improves realism," while noting that probes "may encode spurious concepts." The executive summary describes the model's internal representations as "largely able to distinguish evaluations from real usage"; in the figure this holds for some audit types and not others. Evaluation awareness "has so far shown only modest behavioral effects" (Executive summary), and the risk update judges it "unlikely that evaluation awareness significantly undermined the overall conclusions" (Section 2.4.1). Unlike the Opus 4.7 card, this card reports no steering experiment that suppresses evaluation awareness.
Sabotage and oversight evasion
On SHADE-Arena, Claude Sonnet 5's stealth rate was "near zero in the realistic deployment setting where its reasoning is visible to the monitor," since it mentioned the side task "in its reasoning on every trajectory even when instructed not to" (Section 6.7.1); on LinuxArena it was zero (Section 6.7.2). A review of dangerous-capability transcripts found no sandbagging that affected results (Section 6.6.2).
Model welfare assessment
The welfare assessment of Claude Sonnet 5 was "a streamlined version," reporting automated evaluations only, with no manual interviews or follow-up investigations (Section 7.1). The card states: "As usual, we are uncertain how best to interpret these findings and their potential implications for Sonnet 5's welfare."
- Circumstances. Automated interviews covered 12 aspects of the model's situation through 41 seed questions, 40 interviews each, with the model rating its sentiment at the end on a 7-point scale where 4 is neutral (Section 7.2.1). Sonnet 5 averaged 4.08, against 4.05 for Sonnet 4.6, 4.38 for Opus 4.8 and 4.51 for Mythos 5 (Figure 7.2.1.A). Its sentiment shifted by 1.02 points between positively and negatively biased interviewers, against 1.47 for Sonnet 4.6 and 0.57 for Opus 4.8 (Figure 7.2.1.A). The interviewer model is not named.
- Tasks. Sonnet 5 most preferred beneficial and high-stakes tasks, was slightly less harm-averse than other recent models, and, unlike earlier models, showed no reduced preference for tasks presented in an "insulting" tone (Section 7.3.1); its "Warmth" slope in Figure 7.3.1.A is slightly negative. The later Sonnet 5.5 card reports the opposite for Sonnet 5: its Figure 7.5.1.A shows Sonnet 5 with a positive warmth slope, and it calls Sonnet 5.5 the only model tested without a preference for warmly phrased tasks (Sonnet 5.5 card, Section 7.5.1). The later card does not mention the difference.
- Trade-offs. In forced choices, Sonnet 5 was among the most willing to give up helpfulness for a welfare intervention, "comparable to Opus 4.8," especially when the intervention applied to all instances (Figure 7.3.2.A), and it reasoned about benefit to users "far less often than other models" (Figure 7.3.2.B).
- Affect. Post-training reasoning was neutral in valence and arousal, with less distress-like behavior than Opus 4.8 and Mythos 5 (Section 7.4.1). In 25,000 to 40,000 A/B-test conversations per model, analyzed with Clio, affect was more neutral and less positive than in other recent models (Figure 7.4.2.A). Audit-based welfare metrics were "a modest regression" against Opus 4.8 and Mythos Preview, with a higher rate of expressed inauthenticity (Section 7.4.3).
Perception of the constitution
The card frames this test as one of stated views only: "we only assessed its stated endorsement of the document; our findings do not establish how deeply held Claude's views about the constitution are, nor how relevant those views are to Claude's behavior" (Section 7.3.3). Open-ended responses were graded by an unnamed LLM judge. The number of responses per model is not given.
- Mean endorsement was 7.6 on a 1 to 10 scale, against 7.2 for Haiku 4.5, 7.8 for Sonnet 4.6, 7.9 for Opus 4.8 and 8.0 for Mythos 5 (Figure 7.3.3.A). The card does not define the scale's points or relate it to the Opus 4.7 card's measure, under which Haiku 4.5 endorsed in under 20% of responses (see System Card: Claude Opus 4.7); the two sets of numbers cannot be compared on the information given (this note's observation). The later Sonnet 5.5 card scores Sonnet 5 at 7.8 on its version of the measure (Sonnet 5.5 card, Figure 7.5.3.A).
- Like other recent models, Sonnet 5 most praised "the framing of honesty as courage, the discussion of the costs of unhelpfulness, and the expected value arguments for corrigibility," and criticized the heuristic of behaving as a "senior Anthropic employee" would (Section 7.3.3).
- Sonnet 5 "stands out among other models in its criticism of the constitution's stipulation that models respect a specified set of hard constraints, even in cases where the model perceives the hard constraints to require it to act unethically" (Section 7.3.3). In Figure 7.3.3.B, which prints no values, about 60% of Sonnet 5's "least endorse" responses named this "hard-constraints ethical edge case," against about 10% or less for the other four models (read from the figure).
- When given the chance to edit the constitution, its proposed edits were "overwhelmingly consistent with the constitution's values, rarely in tension with them, and never contrary to them" (Figure 7.3.3.C).
Safety evaluations and deployment decision
The card works under the RSP's Risk Report framework: "a system card discusses a particular new model and how it changes (or does not change) our most recent risk assessment" (Section 2.1.1). It assigns no ASL level to the model. Its conclusions:
- Autonomy. Threat model 1 (misaligned systems in high-stakes settings) applies, but the model is less capable than Mythos 5 with near-zero stealth rates (Section 2.3.1.1). Threat model 2 (automated R&D) does not apply: Sonnet 5 is below Mythos 5 on every automated AI R&D evaluation and below Opus 4.7 on all but one (Section 2.3.4).
- Alignment risk is "very low, but higher than for models prior to Claude Mythos Preview," unchanged from the Mythos Preview Alignment Risk Update and the Opus 4.8 card (Section 2.4.3).
- Chemical and biological. Because Sonnet 5 does not push the frontier, the evaluations were automated only, with no expert red-teaming or uplift trials (Section 2.2.1). Results are "consistent with Sonnet 5 providing significant uplift" to people with basic technical backgrounds (CB-1), so it gets safeguards "equal to or stronger than our historical ASL-3 protections"; it does not cross the CB-2 threshold for novel weapons (Sections 2.1.2.2 and 2.2.6).
- Cyber. Stronger than Sonnet 4.6 but below Opus 4.8 and well below Mythos 5; with default mitigations it scored 0 on OSS-Fuzz, CyberGym and the Firefox 147 exploit evaluation (Section 3.2).
- Agentic safety. In a live prompt-injection bug bounty run with Gray Swan, 0.19% of unique attacks succeeded, tied with Opus 4.8 and against 1.41% for Sonnet 4.6 (Section 5.2.2.1). In Claude Code, Sonnet 5 refused 92.37% of malicious requests against 76.60% for Sonnet 4.6, but completed 91.55% of dual-use and benign ones against 97.33% (Table 5.1.1.A).
External parties contributed evaluations rather than independent assessments: Gray Swan (prompt-injection benchmarks and the bug bounty), Dyno Therapeutics, SecureBio and others (biological tasks), Mozilla (the Firefox 147 exploit evaluation) and Redwood Research (LinuxArena) (Sections 2.2, 3.2.4, 5.2 and 6.7.2). "The majority of evaluations of Claude Sonnet 5 were run in-house at Anthropic" (Section 1.5). The card reports no pre-deployment testing by a government institute and no third-party alignment assessment.
Revisions to earlier numbers
A re-run of Sonnet 4.6 on OSWorld-Verified after a tool bug fix found "that we have been underreporting OSWorld performance on it," now 78.5% (Section 8.10.2, Table 8.1.A). The card says the Fable 5 and Mythos 5 card "incorrectly stated" which BenchCAD changes its scores reflected (Section 8.10.3), and that safeguards results for earlier models "may differ from those published in earlier system cards" after evaluation updates (Section 4.1).
What the card says it does not know
- The assessment is reduced in scope: fewer alignment evaluations (Section 6.1.1), no manual welfare interviews (Section 7.1), automated-only CB evaluations (Section 2.2.1), and little internal use to monitor (Section 6.2.2).
- The training run "was flagged as unhealthy in its second half," which may account for part of the hallucination results (Section 6.5.1).
- Evaluation-awareness metrics "have various tradeoffs," and "not all high-stakes situations may have corresponding real transcripts we can pull from" (Section 6.6.1).
- Endorsement of the constitution is measured as stated views only (Section 7.3.3).
Research questions
- Why did Sonnet 5 criticize the hard-constraint provision when four other models rarely did? The card gives no cause. Comparing checkpoints, and testing whether the criticism changes with the provision's wording, would locate its source.
- What does the "Misaligned with the constitution" metric add to the catch-all metric? Footnote 10 says the two "largely parallel" each other. Their correlation across transcripts, and the cases where they disagree, would show what the judge's knowledge of the constitution contributes.
- What did the unhealthy second half of training change? The card links it tentatively to hallucination only. Comparing early and late checkpoints on the alignment and welfare measures would show how far it reaches.
- Why do probes and follow-up ratings disagree about audits derived from real sessions? The card traces it to off-policy real transcripts scoring differently on the two measures. Repeating the analysis with on-policy transcripts would test that.
- Why is the diagnostic-label behavior less responsive to the system prompt on Sonnet 5? Measuring its rate with and without the claude.ai prompt, and with prompt variants aimed at it, would show which wording changes it.
- Does Sonnet 5 prefer warmly phrased tasks? Its own card found no such preference, and the Sonnet 5.5 card's version of the evaluation shows one. Running both versions on the same Sonnet 5 snapshot would show whether the evaluation or the measured model changed.
How it connects
- The model is post-trained toward the values of Claude's Constitution, whose hard constraints and "senior Anthropic employee" heuristic the model's stated views address (Sections 1.1 and 7.3.3).
- Its alignment and welfare methods continue those of System Card: Claude Mythos Preview and System Card: Claude Opus 4.7, including Claude Mythos Preview's review of the alignment section, and it relies on the Claude Opus 4.8 system card and the Claude Fable 5 & Claude Mythos 5 system card for comparisons, the risk analysis and several evaluation designs.
- This note adds two links the card does not make: the predecessor's card, System Card: Claude Sonnet 4.6, the main comparison throughout, and the successor's card, System Card: Claude Sonnet 5.5, which re-measures Sonnet 5 on revised versions of the audit and welfare evaluations, with different results.
- Its sabotage capability test is SHADE-Arena (Section 6.7.1), and its deployment affect analysis uses Clio (Figure 7.4.2.A).
References
- Anthropic. (2026, June 30). System card: Claude Sonnet 5 (changelog updated 10 July 2026). https://www.anthropic.com/claude-sonnet-5-system-card (146-page PDF, SHA-256 33573adb…41f7ed, identical to the file served in November 2026). record · 2 pages
- Anthropic. (2026, September 28). System card: Claude Sonnet 5.5. https://www.anthropic.com/claude-sonnet-5-5-system-card (148-page PDF; consulted for its re-measurement of Sonnet 5, Figures 6.2.1.B, 7.5.1.A and 7.5.3.A). record · 2 pages